Skip to content
sdq
Back to the libraryGlossary

Glossary

Plain-language definitions of every term used in the lessons.

Every underlined term in a lesson links here. Definitions are kept short and jargon-free.

3-D Secure
The extra check for online card payments: your bank steps in during checkout and asks you to prove it's you, usually with an OTP or an approval in its app.
A/B test
Showing two versions of something, A and B, to two random groups of people at the same time, and measuring which works better before choosing one for everyone.
A2A
Agent-to-agent protocol: an open standard that lets AI agents built by different teams or companies find each other and pass tasks back and forth.
ABAC
Attribute-based access control: each decision uses facts about the person, the item and the situation, such as department, document label or time of day.
Abstraction layer
A thin layer of your own code between your app and outside providers or frameworks, so you can swap one for another without rewriting the app.
Acceptance rate
In speculative decoding, the share of guessed tokens the big model agrees with and keeps. The higher it is, the bigger the speed-up.
Acceptance rate (suggestions)
The share of AI suggestions shown to people that they actually keep, for example by pressing Tab. It is the main everyday measure of a code assistant's usefulness.
Access control
The rules about who may see or change each file or record, and the checks that enforce them every time someone asks.
Access token
A short-lived key that lets an app do a few named things on your behalf, such as read your profile, for about an hour. It is not proof of who you are.
Accessibility tree
A structured outline of everything on a page or screen, such as buttons, links, fields and their labels, that screen readers and agents can read instead of pixels.
Account enumeration
Using a site's own sign-in or reset page to find out which email addresses or phone numbers have accounts there, for example because it says "No account with this email".
Account takeover
When someone other than the owner gets into an account, usually with a stolen or leaked password, and uses it as if it were theirs.
Accuracy
The share of all a model's decisions that were right. It is misleading when the thing you look for is rare, because always saying 'no' already scores very high.
Acknowledgement
A short "got it" reply that confirms a message arrived. Without one, the sender can't be sure, so it keeps trying.
Acquirer bank
The bank that works for the shop or toll plaza receiving the money. It sends payment requests into the network on their behalf.
Activations
The numbers flowing through a model's layers while it works on a particular input, as opposed to the fixed weights. They change with every request.
Active learning
Letting a model pick which unlabelled examples people should label next, usually the ones it is least sure about, so fewer labels teach it more.
Active-active
Running an app in two or more regions that all serve users at the same time.
Adapter
The small file of extra weights that LoRA and similar methods produce. One base model can serve many adapters, one per customer or task.
Adaptive bitrate
The player switching between video qualities on the fly: sharper when your internet is fast, lower quality instead of buffering when it slows down.
Admission control
Deciding at the front door whether to accept a new request, based on how busy the system is, so it never takes on more work than it can finish in time.
Adversarial input
An input deliberately changed by someone, often in ways people barely notice, so that a model gets it wrong, such as a border or a strip of text that hides a banned item.
Agent card
A small public file that describes an AI agent for other agents: what it can do, how to send it a task and what sign-in it needs. Used by the A2A standard.
Agent harness
The code around a model that runs the agent loop: building each turn's context, calling tools, trimming their output, checking work and deciding when to stop.
Agentic RAG
RAG where the model decides what to search for, judges the results, and searches again or elsewhere until it has enough to answer.
Aggregation
Combining many individual records into one summary, like averaging the speeds of 40 phones into one speed for a road.
AI agent
An app that lets an AI model choose its own next steps, such as searching, checking or booking, by asking for tools in a loop until a task is done.
AI framework
A ready-made code library that handles the repetitive work around AI model calls, such as talking to different providers, filling in prompts, calling tools, searching documents and keeping track of state.
AI gateway
A single service between an app and all the AI models it uses. It holds the keys, picks the model, applies limits, retries or switches on failure, and logs every call and its cost.
AI inventory
A company's list of every place it uses AI, with the model and provider, the data involved, the owner and the risk level, kept up to date as uses change.
Alert fatigue
What happens when people get so many alerts, most of them unimportant, that they start ignoring or silencing them and miss the ones that matter.
Alignment tax
The loss of everyday skill a model can suffer when it is tuned hard for safety or a particular style, such as refusing harmless questions or doing worse at ordinary tasks.
All-reduce
A group network operation in which every GPU shares its numbers and every GPU ends up holding the same combined result, such as the average gradient across all of them.
Allotment
A fixed number of rooms or seats set aside for one particular seller, such as three rooms a night that only one travel site may sell. Other sites can't sell those, so they can't clash.
Allow-list
A list of the only things permitted, such as the websites an agent may visit. Anything not on it is refused.
Ambient scribe
An AI note-taker that listens to a conversation, such as a doctor's consultation, and writes a draft record of it for a person to check and sign.
Anchoring
People leaning too heavily on a suggestion they were shown first, such as accepting a model's pre-label they would not have chosen on their own.
Anomaly detection
Automatically spotting numbers that break their usual pattern, such as a sudden jump in spending or errors.
Anonymisation
Removing anything that identifies a person, such as a name or phone number, so data can be used without knowing whose it is.
ANPR (number-plate camera)
Automatic number-plate recognition: a camera that reads a vehicle's number plate from a photo, so the vehicle can be identified without stopping.
Answer relevance
Whether an AI answer actually addresses the question that was asked, rather than a nearby one. An answer can be true to its sources and still miss the point.
Anti-entropy
A background job that regularly compares copies of the data and fixes any differences, so copies that missed changes eventually catch up even if nobody reads them.
API
An Application Programming Interface: an agreed set of requests one program can send another, and the responses it will get.
API gateway
The front door of an app made of many services: it checks each request (login, limits) and sends it to the right service.
API key
A secret code an app sends with its requests to prove who it is to another service, such as a model provider. Anyone holding it can run up charges, so it must be kept private.
Approval fatigue
What happens when people are asked to approve too many things: they start clicking yes without reading, so the risky requests slip through with the harmless ones.
Approximate count
A number that is close enough rather than exact (like "1.2M likes"), used when perfect counting would be too slow.
Approximate nearest-neighbour index
A shortcut structure in a vector database that finds very close matches without comparing against every stored item. It may rarely miss one, in exchange for answering in milliseconds.
Arena leaderboard
A ranking built from many people comparing two anonymous models' answers side by side and voting for the better one.
ARN
Acquirer Reference Number: a 23-digit number given to a card payment or refund as it passes between banks. Your bank can use it to trace where a refund is.
Asynchronous
Done later, without making the caller wait for the result.
At least once
A delivery promise: every message arrives, but some may arrive twice, so receivers must handle repeats.
At most once
A delivery promise: a message is sent once and never retried, so it can't arrive twice, but it may be lost.
Atomic operation
A change that happens as one indivisible step, so nothing else can sneak in halfway. Checking a counter and lowering it in one atomic step stops two servers from both using the last token.
Atomic swap
Switching every reader from an old set of data to a new one in a single step, usually by changing one small pointer, so nobody ever sees a half-finished mix of the two.
Attention
The step in a transformer where each token looks at the other tokens and decides how much each one matters to its own meaning. In 'river bank', 'bank' pays close attention to 'river'.
Attention head
One of many attention units that run side by side in each layer of a transformer. Each learns to look for a different kind of link between tokens, such as which noun a pronoun points to.
Audit log
A permanent, time-stamped record of who did what and when, including actions taken by AI agents, kept so it can be checked later.
Authentication
Checking who someone is, for example by verifying their password, OTP or login token.
Authorization code
A one-time code the sign-in provider hands back to an app after you agree. It works once, for a few minutes, and only together with the app's own secret.
AutoGen
An open-source Microsoft framework in which several AI agents solve a task by talking to each other. It is now mostly in maintenance, with new work moving to the Microsoft Agent Framework.
Automatic prompt optimisation
Using software to search for better prompt wording and examples, keeping whichever version scores best on a test set.
Automation bias
The habit of trusting a computer's suggestion too much, so a person asked to check it ends up agreeing almost every time without really looking.
Autoscaling
Adding servers automatically when traffic rises and removing them when it falls, so you pay only for what you need.
Availability
Whether the app answers at all. "Highly available" means it keeps responding even when some parts fail.
Availability zone
A separate building, or group of buildings, within one cloud region, with its own power, cooling and network. Spreading copies across zones means one building's trouble can't take them all out.
B-tree
The sorted, shallow tree most database indexes use. Each step down narrows the search to a smaller range, so a few steps find one entry among crores.
Back-of-the-envelope estimate
A quick, rounded calculation, like 1 crore requests a day ≈ 100 a second, used to decide which parts of a design matter. Close enough beats exact.
Backchannel
A short sound like "haan", "ok" or "uh-huh" that a listener makes to show they're following. It isn't an interruption, so a voice agent should keep talking through it.
Backend for frontend
A server endpoint shaped for one screen or one app, which gathers data from several internal services and returns it in a single answer.
Backfill
Computing values for past dates that were never recorded, such as working out a new feature for every order in the last year so a model can be trained on it.
Background upload
An upload that keeps going while you use other apps or lock your phone, so you don't have to wait on the screen for it to finish.
Backpressure
Slowing down or rejecting new work when a downstream part can't keep up, instead of letting work pile up.
Backup
A copy of your data kept somewhere else, so you can get it back if the original is lost or broken.
Backward compatibility
When a new version of a system keeps working with older versions of the things that talk to it, like a server that still understands requests from last year's app.
Bandwidth
How much data is sent over the network. Cloud providers charge for data sent out to viewers, so popular pictures and videos can cost more to send than to store.
Barge-in
Letting you interrupt an automated voice. As soon as you speak, the system stops talking and listens.
Base 62
Writing numbers using 62 characters (0–9, a–z and A–Z) instead of just the 10 digits, so big numbers fit in very few characters. A number in the hundreds of crores takes only six.
Base model
A model straight after pretraining, before it is taught to follow instructions. Also the unchanged model that fine-tuned add-ons are attached to.
Base rate
How common something is in the first place, such as 1 fraud in every 1,000 payments. When it is rare, even a small false alarm rate produces more false alarms than real catches.
Baseline model
The simplest reasonable way to make a prediction, such as today's rules or a very simple model. Any new model has to beat it to be worth its cost.
Batch API
A way to send non-urgent model requests in bulk to be answered within hours, usually at a lower price than instant answers.
Batch job
A large piece of work run in one go on a schedule, often overnight, that goes through lots of stored data and saves the results for later.
Batch scoring
Running a model over a large list of users or items in one scheduled job, often overnight, and saving every answer so the app can simply look it up later.
Batching
Collecting many small things and handling them together, like adding up 1,000 likes in one go instead of one by one.
Beamforming
Using several microphones a little apart to listen mainly in one direction. Sound from that direction reaches each microphone at slightly different moments, so lining those moments up makes it louder while sounds from elsewhere are turned down.
Behavioural interview
A job interview round made of "tell me about a time…" questions, where you describe real situations from your past work so the interviewer can judge how you think and act.
Benchmark
A standard public test used to compare AI models on one kind of task, such as maths problems or coding.
Benchmark contamination
When a benchmark's questions or answers leaked into a model's training data, so its score overstates its real ability.
Benchmark saturation
When the best AI models all score close to full marks on a test, so differences between them are mostly noise and the test can no longer tell them apart.
Best-of-N
Asking a model for several versions of an answer, such as 8, and letting a scorer, often another model, pick the best one. The pick can only be as good as the scorer.
Biometric
Something about your body that identifies you, such as a fingerprint, the pattern of your iris or your face, used in place of a password or ID card.
Bit
The smallest unit of data, a single 0 or 1. Eight bits make one byte. Network speeds are quoted in bits per second, file sizes in bytes.
Bitrate ladder
The set of qualities a video is converted into, from a small, low-data version to a sharp, high-data one, like rungs on a ladder. The player climbs up or down it as the connection changes.
Blast radius
How much damage one failure, mistake or compromised part can do before something stops it. Good designs shrink it by limiting what each part can reach, for how long and how fast.
Blocklist
A list of things that must be refused, like FASTags with too little balance or unfinished KYC. Systems keep a copy close by so they can check it instantly.
Bloom filter
A tiny, fast structure that answers "definitely not seen" or "probably seen", used to skip work on things already handled.
Blue-green deployment
Running the old and new versions side by side, switching traffic to the new one once it checks out, and keeping the old one ready to switch back to.
BM25
The most common way keyword search scores a passage. Rare words count more than common ones, repeats help a little less each time, and long passages are marked down slightly.
Bot
A program pretending to be a person, often used to grab tickets or deals faster than humans can.
Bottleneck
The part of a system that limits everything else, like the narrow neck of a bottle.
Bounding box
The rectangle on a page or image that marks where a word, block or object sits, given as its position and size. It lets a system highlight, cite or hide that exact spot.
Branch
A separate line of changes in a code project, so work can happen without touching the main version until it has been reviewed and merged.
Broker
One server in an event-streaming cluster. Brokers store the partitions, accept writes and serve reads; a cluster usually has several so it survives one failing.
Browser automation script
A program that drives a web browser through fixed, pre-written steps, such as 'click the element with this name, then type here'. Fast and cheap, but it breaks when the page's design changes. Selenium and Playwright are well-known tools for writing them.
Brute force
Trying every possible answer, one after another, until one works. Limits on the number of tries stop it.
Bucket
A named container in object storage that holds files, such as one bucket for photos and another for invoices. Access rules are usually set per bucket.
Buffer
Video downloaded ahead of the part you're watching, so short internet slowdowns don't stop playback. "Buffering" means it ran out.
Bulk load
Writing a very large set of new values into a database in one planned run, such as a crore saved lists every night, instead of one change at a time as users act.
Bulkhead
Giving different kinds of work their own separate capacity, like the sealed compartments in a ship's hull, so a flood of one kind can't use up everything and sink the rest.
Burn rate
How fast an error budget is being used up compared with the steady pace that would last exactly the whole period. A burn rate of 10 empties a 30-day budget in 3 days.
Byte
A unit of data made of 8 bits. Memory and files are measured in bytes (GB), network speeds usually in bits (Gbps), so divide a network speed by 8 to compare them.
Cache
A fast store that keeps copies of frequently used data so you don't have to fetch it from the slow source every time.
Cache eviction
Deciding what to remove from a full cache to make room, often the item used least recently. Removed items can be fetched or rebuilt later, just more slowly.
Cache hit
A request the cache could answer from its own copy.
Cache invalidation
Removing or marking out of date the cached copies that depend on something that has just changed, so nobody is served the old version.
Cache key
The label a cache uses to find a stored answer. Everything that can change the answer, such as the question, the user's language or the prompt version, has to be part of it.
Cache miss
A request the cache couldn't answer, so it had to go to the original source.
Cache stampede
Many requests missing the cache at the same moment, often because a popular entry has just expired, so they all rush to the slow, costly source together.
Cache thrashing
When a cache is too small for what is being asked of it, so it keeps throwing out items that are needed again moments later. Almost every request becomes a slow miss.
Cache-aside
The most common way to use a cache: the app checks the cache first, and on a miss reads the database and saves a copy in the cache for next time.
Cache-aware routing
Sending each request to the server that already holds the cached notes for its opening, or for that user's chat, instead of spreading requests evenly. It keeps cache hit rates high across many servers.
Calibration
Adjusting a model's scores so they match how often things really happen, so a score of 0.7 means about 70 in 100 such cases turn out bad. It keeps a threshold meaningful across model versions.
Calibration set
A small sample of typical inputs run through a model while compressing it, to see which numbers matter most so they can be rounded with the least harm.
Canary release
Sending a small share of real traffic to a new version first, and only rolling it out further if it stays healthy.
Canary token
A secret marker, such as a random string, hidden in a system prompt or a file. If it ever shows up in an output or outside, you know that content has leaked.
Candidate
A post that could go into your feed, gathered before the app decides which ones to actually show you.
Candidate generation
The first, fast step of a recommender: pulling a few hundred possible items out of a huge catalogue, before a slower model decides which to show.
CAP theorem
The rule that when the network between machines breaks, a distributed system must choose between staying available (answering, but maybe with disagreeing data) and staying consistent (refusing some requests).
CAPTCHA
A small puzzle, like typing wobbly letters, that is easy for people and hard for bots.
Caption cue
One caption on screen: a line or two of text with a start time and an end time.
Card chip
The small gold chip on a card. It's a tiny computer holding a secret key that can't be copied off, used to prove the real card is there.
Card network
The company that connects card-holders' banks with shops' banks and carries card payments and refunds between them, such as Visa, Mastercard or RuPay.
Card token
A stand-in number a website saves instead of your real card number. It only works for payments on that site, so if it leaks it's useless anywhere else.
Cardinality (of metrics)
How many separate series a metric splits into, one for every combination of its labels. Labelling by city gives dozens; labelling by user ID gives crores and can overwhelm the monitoring system.
Cascading failure
When one part's failure spreads to the parts that depend on it, which then fail and spread it further, until a small fault becomes a big outage.
Catastrophic forgetting
When fine-tuning on a narrow task makes a model lose skills it had before, such as general knowledge or following other kinds of instruction.
Causal consistency
A promise that things which depend on each other are seen in the right order, so nobody sees a reply before the message it answers.
CDN
Content Delivery Network: many caches spread around the world that serve files from a location near each user.
Chain of thought
Asking a model to write out its working before giving the answer, which helps on problems with several steps.
Change data capture
Turning every change in a database into an event, so other systems can follow along without the app writing twice.
Channel manager
Software a hotel uses to sell its rooms on many websites at once. It keeps every site's room counts and prices up to date and passes their bookings back to the hotel.
Chargeback
A card payment reversed by the customer's bank after a dispute, such as 'I didn't make this purchase'. Confirmed chargebacks teach fraud models what fraud looks like.
Checkpoint
A saved record of a task's progress after a step, so it can resume from there after a crash instead of starting over.
Checksum
A short value worked out from some data so that a change or a typo shows up: a fingerprint of a file, or the last check digit of a card or Aadhaar number.
Chunk
A small piece of a longer document, a few paragraphs at most, stored and searched on its own so an AI assistant can fetch just the part it needs.
Chunked prefill
Reading a long prompt in small pieces mixed into the normal steps for other chats, instead of all at once, so their answers keep flowing while it is read.
CI/CD
Continuous integration and continuous delivery: every change is tested automatically and, if it passes, released through a fixed pipeline of checks.
Circuit breaker
A guard around calls to another service: after too many failures it stops calling for a while and fails fast instead, like a tripped MCB.
Citation
A link from part of an AI answer to the exact passage or page it came from, so people can check it.
Claim with a time limit
When a worker takes an item from a queue, it is locked to them only for a few minutes. If they don't finish in time, the lock runs out and the item goes back for someone else.
Class imbalance
When one outcome is far rarer than the other in the data, such as 1 fraud in 5,000 payments. Plain accuracy becomes misleading, and training often samples the common class down.
Classifier
A model that sorts an input into one of a few categories, such as spam or not spam. Usually small, fast and cheap compared with a language model.
Click-through rate
The share of times something was shown that it got tapped or clicked. If an item is shown 1,000 times and tapped 50 times, its click-through rate is 5%.
Client
Whatever sends requests: a browser, a mobile app, or another service.
Client secret
A password that belongs to the app itself, given to it by the sign-in provider and kept on the app's server, so the provider knows it's talking to the real app.
Clock skew
The difference between two computers' clocks. Even synced clocks drift apart by milliseconds or more.
Closed model
A model whose weights are not published, so it can only be used through its maker's app or API.
Co-purchase
Two items bought in the same order. Counting how often each pair is bought together across many orders gives "frequently bought together" lists.
Code-mixing
Using two languages in one sentence, like 'timer lagao, and play some music'.
Codebase index
A searchable map of a code repository, its files, functions and how they connect, that lets an AI tool quickly find the code relevant to a task.
Codec
A method for squeezing video into a small file and unpacking it again for playback, such as H.264 or AV1. Newer codecs make smaller files, but older phones may not be able to play them.
Coding agent
An AI agent that works on a codebase: searching and reading files, editing them, running commands and tests, and repeating until the task is done.
ColBERT
The best-known late-interaction search model. It scores a passage by matching each question token with its closest partner among the passage's tokens.
Cold start
When a new product or new user has no history yet, so suggestions based on past behaviour have nothing to work with.
Cold storage
A cheaper kind of storage for files that are rarely opened. Keeping files there costs less, but fetching them is slower or costs a little each time.
Collaborative filtering
Suggesting things to you based on what shoppers or viewers with tastes similar to yours chose. Many people's choices together pick for each person.
Collapse key
A group name attached to related notifications, such as one order's updates, so a newer one replaces the older one instead of piling up.
Collect request
A UPI request in which someone asks you to pay them. The money moves only if you approve it with your UPI PIN, so it always means paying, never receiving.
Colour profile
A tag inside an image file that says how its numbers map to real colours. The same numbers can look duller or brighter depending on the profile.
ColPali
A family of search models that treat each document page as a picture. They keep a small vector for every patch of the page and match the question's words against those patches, with no text extraction needed.
Community (graph)
A cluster of entities in a knowledge graph that are more tightly linked to each other than to the rest, found by a graph algorithm. GraphRAG writes a short summary for each one.
Compaction
Replacing the older part of a long conversation or agent run with a short summary, so the work can carry on within the context window.
Compaction (in a database)
Background work in an LSM-tree database that merges several sorted files into fewer, bigger ones, keeping only the newest version of each key and dropping deleted ones.
Compare-and-set
A single step that changes a value only if it still holds what you expect, like "mark this seat held, but only if it is still free". If someone changed it first, nothing happens and you are told.
Compensating action
A step that undoes an earlier one, like a refund for a payment or cancelling a booking.
Composite index
An index on two or more columns, sorted by the first and then by the next within it, like a phone directory sorted by surname and then first name. The column order decides which questions it can speed up.
Compression
Making a file smaller so it's quicker to send and cheaper to store, usually by dropping detail you won't notice, like shrinking a video to suit a phone screen.
Compute
The amount of calculation, done on GPUs or other chips, used to train or run a model. More compute usually means higher cost and, up to a point, better results.
Computer-use agent
An AI agent that works a website or app the way a person does, by looking at the screen and then clicking or typing.
Concept drift
When the link between inputs and the right answer changes, so the same-looking case now means something different. Fraudsters changing their tricks is the classic example.
Concurrency
Many things happening at the same moment, like thousands of people pressing "Book" at exactly 10:00:00.
Conditional request
A request that says "send this only if it has changed since a given time or version". If nothing changed, the server replies with a tiny "not modified" answer instead of the whole page.
Confidence calibration
How well a confidence score matches reality. A well-calibrated model that says it is 90% sure is right about 90 times in 100; a badly calibrated one may say 90% and be right only 60 times.
Confidence score
A number from a model saying how sure it is about one result, used to decide when to trust it and when to send it to a person.
Confirmation prediction
A guess, based on how waitlists moved on similar trains and dates in the past, of how likely a waitlisted ticket is to get a berth. It's a forecast, not a promise.
Confused deputy
A helper program with wide powers that is tricked into using them for someone who has fewer rights, such as an assistant opening a file for a user who couldn't open it.
Congestion control
TCP's habit of starting slowly, speeding up while everything arrives, and slowing down when packets go missing, so many connections can share a busy network without jamming it.
Connection pool
A fixed set of open connections or worker slots that a service keeps ready for calls to another service. When all of them are in use, new calls have to wait.
Connector
A small program that links one outside system, such as a wiki, shared drive or ticket tool, to an AI app: it fetches documents, their changes and who may read them.
Consensus
A way for a group of machines to agree on one answer, such as who the leader is or what order writes happened in, even if some of them fail. Raft and Paxos are well-known methods.
Consensus label
The final answer for an item that several labellers marked, usually the majority vote, with a senior reviewer deciding when they split.
Consistent hashing
A way to assign keys to servers so that adding or removing a server moves only a small share of the keys.
Constitutional AI
A training method in which a model judges and rewrites answers against a written list of principles, so AI feedback replaces much of the human comparison work. People write the principles and check samples.
Constrained decoding
Blocking, at each step of writing, every token that would break a required format such as a schema, so the model can only produce replies of the right shape.
Consumer group
A set of readers that share the work of reading an event stream, each group keeping its own position in it.
Consumer lag
How far a reader is behind the newest event in a stream, counted in events or in seconds. Growing lag means the reader can't keep up.
Consumer lag
How far behind a stream-processing job is: how many events, or how many minutes of events, are waiting that it has not handled yet.
Contactless payment
Paying by holding a card or phone near the card machine instead of inserting it. The two talk for a moment over short-range radio (NFC).
Content credentials
A signed note attached to a picture or video saying how it was made, for example by which AI tool and when. It can be lost if the file is screenshotted or re-saved.
Content moderation
Checking text, pictures or video for harmful or banned material before it's shown to others, usually by automatic checks, with people reviewing the unclear cases.
Context engineering
Deciding what goes into a model's context on every call, in what order and form, so it has exactly what it needs and nothing that distracts it.
Context rot
Answer quality slipping as a context fills with old, irrelevant or conflicting material, even before it reaches the window's limit.
Context window
Everything an AI model reads on one call: the app's instructions, the chat so far, any documents and the question. It has a maximum size, counted in tokens.
Contextual bandit
A bandit that uses information about the person and the moment (tastes, device, time of day) to choose, so it learns the best option for each kind of situation rather than one winner for everyone.
Contextual retrieval
Adding a short line to each chunk before it is stored, saying which document and section it comes from, so a chunk that makes little sense alone can still be found.
Continued pretraining
Training an existing model further on lots of raw, unlabelled text from one field, such as medical notes or legal filings, so it learns that field's vocabulary before it is tuned on example tasks.
Continuous batching
Letting requests join and leave a running batch between token steps, so the GPU stays busy and nobody waits for the slowest request in their group.
Contrastive learning
Training a model by showing it pairs that belong together, like a photo and its caption, and pairs that don't, so it learns to place matching pairs close together and the rest far apart.
Convolutional neural network (CNN)
A kind of image model that slides many small pattern detectors across a photo, finding edges first, then textures, then shapes like a collar or a shoe sole, layer by layer.
Coordination service
A small, highly reliable store that a group of machines uses to agree on things like who the leader is, holding locks and watching for changes. etcd and ZooKeeper are well-known examples.
Copy-on-write
Letting many users share one copy of some data, and making a private copy for a user only at the moment they need to change it.
Correlated errors
Mistakes that several models or checkers make together, usually because they learned from similar data. When errors are correlated, voting and cross-checking add cost without catching them.
Cosine similarity
A common score for how alike two embeddings are, based on the angle between them. A score near 1 means very similar in meaning.
Cost attribution
Tagging every model call with the feature, team or customer it served, so spending can be traced to its source.
Cost per task
What it really costs to get one job done with an AI model, counting retries, extra tokens and any hand-offs to people, rather than just the price per token.
Counter
A stored number that goes up or down, like the number of likes on a post.
CQRS
Command Query Responsibility Segregation: handling changes and reads with separate models, with reads served from views built for each screen.
Crash report
A short message an app sends automatically when it crashes or freezes, saying what went wrong, on which phone model and which app version.
crash-free users
The share of people using an app version who had no crash at all in a period, such as a day. At 99.5%, 1 person in 200 saw at least one crash.
Crawler
A program that visits web pages, saves what's on them and follows their links to find more pages. Search engines run huge numbers of them all the time.
Crawler trap
A set of web pages that keeps producing new links forever, like a calendar with an endless "next month" link. A crawler without limits would never finish it.
CRDT
Conflict-free replicated data type: a way of storing shared data so that copies can accept changes separately and always merge to the same result, in any order, without a central server deciding.
Credential stuffing
Taking email and password pairs leaked from one site and trying them on many other sites, hoping people reused the same password.
CrewAI
An open-source framework for teams of AI agents, each given a role such as researcher or writer, working through tasks one after another or under a manager agent.
Cross-border transfer
Sending personal data to be stored or processed in another country, for example calling an AI model hosted abroad with a customer's chat.
Cross-encoder
A model that reads a question and one passage together and scores how well the passage answers it. It is far more careful than comparing stored embeddings, but too slow to run on a whole library.
Crypto-shredding
Deleting data by destroying the encryption key it was locked with. Every copy, including old backups, becomes unreadable at once.
Cryptogram
A one-time code a card's chip makes for each payment from its secret key, the amount and a counter. The bank can check it, and a copied one is useless.
CSAT
Customer satisfaction score: the share of customers who rate a support chat as good, usually from a quick rating at the end.
Cursor
A bookmark the server hands back with each page of results, marking exactly where that page ended, so the next request carries on from there without repeats or gaps.
Custom vocabulary
A list of names and special words given to a speech-to-text model so it recognises them, such as players' names and cricket terms like "doosra".
Cut-off time
The last moment an order can be placed and still go out on today's truck or flight. Orders placed after it wait for the next one.
Daily active users (DAU)
How many different people use an app on a typical day. It's the usual starting point for estimating how much traffic a system must handle.
Dark store
A small warehouse set up like a shop, with shelves and staff but no walk-in customers. It exists only to pack online orders for the neighbourhood around it.
Data augmentation
Making extra training examples by changing existing ones a little (rotating, cropping, brightening, adding blur) so the model learns to ignore those differences.
Data breach
When attackers break into a company's systems and copy its data, such as its list of accounts and password fingerprints.
Data centre
A building full of servers, with its own power and cooling. Big apps run in several, often in different cities.
Data contract
An agreement with the team that supplies some data on which fields, formats and update times they will deliver, checked automatically on every delivery so a surprise change is caught early.
Data drift
When the inputs a model sees change from what it learned on, such as new cities, new payment methods or a different mix of customers, even though the right answers follow the same rules.
Data fiduciary
Under India's data protection law, the organisation that decides why and how personal data is used. It stays responsible even when another company processes the data for it.
Data flywheel
A loop where a product's own use produces the data that makes it better: people use it, their reactions are logged, and those logs become tests and training examples for the next version.
Data labelling
Marking examples with the right answer, such as tagging a support chat as 'refund', to make training and test data. Done by people, or by models checked by people.
Data leakage
Private information, such as another customer's details or internal documents, ending up in an AI answer or log where it shouldn't be.
Data lineage
A record of where each piece of data came from and every step that changed it, so you can trace an answer or a training example back to its source, or remove everything built from one source.
Data loader
The part of a training job that reads, decodes, shuffles and batches training examples and hands them to the GPUs, ideally fast enough that they never wait.
Data localisation
A legal rule that forces certain data to be stored, and sometimes processed, only inside one country, such as India's rule for payment system data.
Data minimisation
Collecting and keeping only the data a feature truly needs, and deleting the rest once it has served its purpose, so there is less to misuse or leak.
Data mix
The shares of each kind of text, such as web pages, code, books and maths, in a model's training data. Changing the mix changes what the model ends up good at.
Data model
The main kinds of records a system stores and the key fields of each, such as an orders table with an order ID, status and rider ID. It's sketched before the boxes so you know what each part must keep.
Data parallelism
Giving every GPU its own copy of the model and a different slice of the training data, then averaging what they all learned after each step so the copies stay identical.
Data pipeline
A chain of automated steps that collects data, cleans it and delivers it where it is needed, on a schedule or as it arrives.
Data poisoning
Slipping bad or misleading examples into data a model learns from or retrieves, so it later behaves the way an attacker wants.
Data principal
Under India's data protection law, the person the personal data is about. They can ask what is held about them, have it corrected, and ask for it to be erased.
Data processor
A company that handles personal data on behalf of the organisation that owns the relationship, such as a cloud host or an AI model provider. The owner stays responsible for what the processor does.
Data residency
A rule that certain data must be stored and processed only in a particular country or region, such as India.
Data warehouse
A large database built for analysing lots of history at once, such as every order from the last two years. Slow for single lookups, good for big reports and training data.
Database
A program that stores data safely and lets you query it, even after restarts and crashes.
Dead-letter queue
A separate holding queue for jobs that keep failing after a few retries, so they stop blocking other work and someone can look at them later.
Deadline propagation
Passing the time by which a request must finish along to every service it calls, so each one knows how long is left and stops work the caller has already given up on.
Debounce
Waiting for a short pause before acting, so a burst of keystrokes sends one request instead of one per key.
Debrief (interview)
The meeting after a set of interviews where the interviewers read each other's written notes and scores and agree on one hiring decision.
Decision threshold
The score at which a system changes what it does, for example allowing low-risk payments, asking for an extra check in the middle band and blocking high-risk ones.
Decode
The second phase of answering, where the model writes its reply one token at a time, each step waiting for the one before. Its speed sets how fast the words stream out.
Decoder-only model
A transformer that reads text left to right and predicts the next token. Chat models are built this way.
Dedicated deployment
Running a model on GPUs that serve only one customer, instead of sharing them. It costs more but gives that customer steady speed and full isolation.
Deduplication
Spotting and dropping repeats, so the same message or action only counts once.
Deep dive
The part of a design interview where you zoom in on one hard piece of the system, such as where the load is heaviest, and work out its details and failure cases.
Deepfake
A fake picture, video or voice recording, made with AI, that shows a real person doing or saying something they never did.
Defence in depth
Protecting a system with several separate layers, such as limited access, a sandbox, a network fence and a person's approval, so that when one layer fails the others still stop the harm.
Deflection rate
The share of support chats that end without a human agent joining. It counts customers who gave up as successes, so it should always be read next to the resolution rate.
Delegated access
An AI app acting on someone's behalf with a token limited to that person's own permissions, rather than with an all-powerful master key.
Delivery promise
The delivery date or time a shop shows before you buy, worked out from where the item is and how long it takes to reach you. The shop tries hard to keep it.
Delivery receipt
A note sent back by an SMS, push or email provider saying a message actually reached the person's phone or inbox, or that it failed.
Delivery zone
The area on the map that one store delivers to, usually a few kilometres around it. Addresses inside are served by that store; addresses outside are not.
Delta update
Sending only what has changed since the last version, like "one more run and a six", instead of the whole thing again. It keeps frequent updates tiny.
Demand forecasting
Predicting how much of each item people will buy, in each area and at each time, from past sales and things like weather and festivals, so the right stock is ready in advance.
Demand spike
A sudden jump in how many people want something at the same time, like everyone leaving a stadium together.
Denial of wallet
An attack or bug that runs up huge bills by making an AI system do expensive work again and again, such as sending endless very long requests.
Dependency
Someone else's code that your app relies on, such as a library or framework. When it changes or breaks, your app can change or break with it.
Deprecation
A warning from the makers of a library or service that a feature will be removed in a later version, usually naming what to use instead. It is a countdown, not just a note.
Deterministic replay
Recovering a crashed run by running its code again from the start while feeding back the saved result of every step already done. It only works if the code makes the same choices each time, so anything variable has to be a recorded step.
Device fingerprint
A set of details about a phone or computer, such as its model and settings, combined to recognise it again even without a login.
Device token
An address for one app on one phone, given out by Google or Apple, that a server uses to send that phone notifications.
Diff
A list of exactly which lines were added, removed or changed between two versions of a file, used to review changes.
Diffusion language model
A language model that writes by starting with a rough draft of many tokens at once and refining it over several passes, instead of one token at a time.
Diffusion model
A kind of AI model that makes pictures by starting from random noise and cleaning it up over many small steps, guided by a text description.
DigiLocker
The Government of India's app for official documents in digital form, such as marksheets, driving licences and Aadhaar. Boards and departments send documents to it directly.
Digital signature
A seal made with a secret key and added to a piece of data. Anyone with the matching public key can check the data is genuine and unchanged, but only the key's owner can make the seal.
Dimensions (of an embedding)
How many numbers are in each embedding. More dimensions can hold finer detail about meaning but take more storage and memory.
Disaster recovery
The plan for surviving the loss of a whole data centre: how much data you might lose (RPO) and how fast you're back (RTO).
Disk-based index
A vector index that keeps small compressed codes in memory and the full vectors and links on fast SSD storage, so huge collections fit on far less memory for a little extra time per search.
Distillation
Training a small AI model on a big model's checked answers for one task, so the small one copies that behaviour at lower cost.
Distributed tracing
Following one request across every service it touches, with how long each step took.
Distributed training
Training one model on many GPUs at once, often across many machines, because it is too big for one GPU's memory or would take years on one GPU.
Distributed transaction
One action that has to update several services or databases, all or nothing, like booking a flight and a hotel together.
DLT registration
India's system where businesses must register their sender name and the exact wording of their SMS templates before mobile networks will deliver their texts. It cuts spam and fake messages.
DNS
The Domain Name System: the internet's phone book, turning names like shop.example into server addresses.
DNS resolver
The computer, usually run by your internet provider or a public service, that looks up website addresses on your behalf and remembers recent answers so the next lookup is quick.
Document database
A NoSQL database that stores each record as a self-contained bundle of fields, often in JSON, so different records can have different fields.
Document-parsing model
A compact AI model trained only to turn page images into structured text, such as headings, paragraphs and tables in reading order, with where each sits on the page. It costs far less than a large general model.
Domain gap
The difference between the data a model learned from and the data it meets in real use, such as clean studio product photos versus dim, blurry phone photos.
domain name
The human-friendly name of a website, like shop.example. Computers can't use it directly, so it gets turned into a number (an IP address) first.
DPDP Act
India's Digital Personal Data Protection Act, which sets rules for collecting and using personal data, including consent, use only for the stated purpose, and erasure on request.
DPO
Direct preference optimisation: training a model straight from pairs of preferred and rejected answers, without building a separate reward model. Simpler and cheaper than RLHF.
Draft heads
A few small extra layers attached to a big model that guess several tokens ahead using the big model's own working. They let it speed up writing without loading a separate draft model.
Draft model
A small, fast AI model that guesses the next few tokens so a big model can check them all at once. The big model has the final say on every token.
Draining
Taking a server out of service gently: it stops getting new requests, finishes the ones already in progress, and only then shuts down.
Drift
AI quality slowly slipping after release because the questions people ask, the data, or the model itself have changed since it was tested.
DRM
Digital rights management: locks on songs, films and books so they only play in the approved app, on your account, for as long as you're allowed.
DSPy
An open-source framework where you describe what each step of an AI program takes in and gives back, and an optimiser writes and tunes the prompts for you.
DTMF (keypad tones)
The tones a phone sends when you press its number keys. Systems read them exactly, so they're the safest way for a caller to enter digits like an account number.
Dual write
Saving a change in one system and then separately telling another system about it, as two steps. A crash between them leaves the two out of step.
Durability
The promise that once the system says your data is saved, it stays saved, even if a machine crashes or a disk fails a moment later.
Durable execution
Running a long, multi-step task so that each finished step is saved, letting it carry on from where it stopped after a crash instead of starting over.
Durable timer
A timer saved in a database rather than held in a running program, such as 'remind the manager in 2 days'. It still fires after servers restart or new code is deployed.
Dynamic batching
Waiting a few milliseconds to gather whichever requests have arrived, then running them as one group. Batches fill better, but the group still runs to the end together.
Dynamic QR
A QR code made fresh for one bill, with the exact amount and an order number inside. A static QR is the same for every customer.
Echo cancellation
Removing a device's own sound from what its microphones hear, so it can hear you while it is talking or playing music.
Edge server
A server close to where you are (often in your own city) that keeps copies of popular content, so it reaches you faster than from a distant data centre.
Edit conflict
When two people change the same part of something at the same time, each without knowing about the other. The system needs a rule to fit both changes together.
Edit distance
How many single-letter changes (adding, removing, swapping or replacing a letter) it takes to turn one word into another. "iphon" is one change away from "iphone".
eKYC
Electronic "know your customer": checking a person's identity online, for example by matching their fingerprint against Aadhaar, instead of with paper forms and photocopies.
Electronic health record
A hospital or clinic's digital file for each patient, holding visits, notes, test results and medicines. Notes become official once a doctor signs them into it.
Embedding
A list of numbers that captures the meaning of a piece of text or a picture, so things with similar meanings get similar numbers.
Embedding model
An AI model that reads text or a picture and returns its embedding, a list of numbers for its meaning. It writes no words, so it is much cheaper and faster than a chat model.
Encoder model
A transformer that reads a whole text at once, in both directions, to understand it rather than to write. Embedding models and many classifiers are encoders.
Encoder-decoder model
A transformer with two halves: an encoder that reads the whole input to understand it, and a decoder that writes the output while looking back at it. Classic translation systems use this shape.
Encrypted backup
A copy of your chats saved to the cloud and locked with a password or key only you know, so the cloud company and the chat app can't open it.
Encryption key
A secret code used to scramble and unscramble data. Without the right key, a scrambled file or message is just gibberish.
End-of-turn detection
Deciding when a speaker has actually finished their turn, not just paused. Small models listen to the tone and the words, which works far better than waiting for a fixed stretch of silence.
End-to-end encryption
Scrambling a message on the sender's phone so only the receiver's phone can unscramble it. Servers in between can pass it on but can't read it.
Endpoint
One entry on an API's menu: a method and an address that does one job, such as POST /orders to place an order or GET /orders/812 to see one.
Ensemble
Combining answers from several model calls or several models, for example by voting, to get a more reliable result than any one alone.
Entity extraction
Using a model to pick out the people, places, organisations, products and other named things in text, and how they relate.
Entity resolution
Working out when different names refer to the same thing, such as 'Shakti Steels' and 'SSPL', and merging them into one record. Mistakes here create links that don't exist.
Enum
A field that may only hold one value from a fixed list, such as food, travel or stay. Short for enumeration.
Episodic memory
Memories of specific past conversations or events, such as 'last week we planned a trip to Goa'.
Epoch
One full pass of training through every example in a training set. Fine-tuning usually runs for a few epochs; too many and the model starts memorising its examples.
Epsilon-greedy
A simple bandit rule: pick the best-known option most of the time, and a random option a small fixed share of the time, such as 5%.
Erasure coding
Splitting a file into pieces and adding a few extra repair pieces, spread across machines, so the file can be rebuilt even if several pieces are lost. It protects data using far less space than full copies.
Erlang C
A classic formula from call centres that works out how long items wait in a queue, given how fast they arrive, how long each takes and how many people are working.
Error analysis
Reading a batch of real AI answers, noting everything wrong with each, and sorting the notes into a few failure types. Those types then become the things an eval scores.
Error budget
The share of requests allowed to fail under an SLO. While it lasts, teams ship freely; when it's gone, they fix reliability.
Error correction
Storing extra, overlapping copies of data so the original can be rebuilt even if part of it is damaged or unreadable.
Escalation rate
The share of requests a cheap first step passes on to a stronger, dearer one, for example 25 in every 100 questions sent on to the large model. A sudden rise usually means something upstream has broken.
ETA
Estimated time of arrival: how long until the driver, the food or the train reaches you.
EU AI Act
The European Union's law on artificial intelligence. It sorts AI uses by risk, bans a few, puts strict duties on high-risk ones and lighter transparency rules on others.
Eval
A test for AI behaviour: a saved set of example inputs, each with what a good answer must contain, used to score a model or prompt before and after changes.
Eval gate
A step in a release pipeline that blocks a change unless its eval scores stay above set thresholds.
Eval metric
A rule that turns an AI answer into a score, such as 'matches the right category' or 'valid and under 50 words', so different prompts or models can be compared by numbers rather than by feel.
Evaluator-optimiser
A pattern where one model produces an answer and another grades it with feedback, repeating until the grade is good enough.
Event
A small record that something happened, like "order placed" or "post liked", that other parts of the system can react to.
Event log
An append-only, ordered record of events that is kept for days or weeks, so many systems can read it independently and even read it again from the start. Apache Kafka is the best-known example.
Event sourcing
Storing every change as an event (like passbook entries) and calculating the current state from them, instead of overwriting it.
Event time
The moment something actually happened, such as when a payment was made, as written inside the event itself. It can be very different from when a system gets round to reading it.
Eventual consistency
A guarantee that copies will agree once updates stop arriving, but may briefly disagree.
Exactly once
The ideal delivery promise: every message arrives once and only once. Networks can't guarantee it on their own, so apps get the same effect from retries plus receivers that ignore repeats.
Exactly-once processing
Making sure each event changes the results once and only once, even when a reader crashes and reads some events again. In practice it is built from replays plus updates that are safe to repeat.
Excessive agency
Giving an AI system more tools, permissions or freedom to act than its job needs, so a mistake or trick can do real harm.
Exfiltration
Quietly sending stolen data out of a system, for example by tricking an AI agent into putting private details inside a link or an email to an outside address.
EXIF data
Hidden details a camera saves inside a photo file, such as the phone model, the time and often the exact location where it was taken. Apps usually remove it before sharing a photo.
EXIF data
Hidden notes a camera saves inside a photo file, such as which way the phone was held, the camera model and the colour settings.
Expand and contract
Changing a database or API in safe steps: first add the new thing alongside the old, then move everything over, and only later remove the old thing, so old and new code both keep working throughout.
Explicit feedback
A reaction a user gives on purpose, such as a thumbs up or down, a star rating or a typed correction. Clear in meaning, but only a small share of users ever give it.
Exploration
Deliberately showing a small share of items the system is less sure about, so it keeps learning what people like instead of only repeating its past guesses.
Exponential backoff
Waiting longer after each failed try (say 1 s, then 2 s, then 4 s) so a struggling server gets time to recover.
Face clustering
Grouping photos of faces by how close their numbers are, so each group is probably one person. It doesn't know names until you add them.
Face match
Comparing two face pictures, such as a selfie and the photo on an ID card, and scoring how likely they show the same person.
Fail closed
Blocking things when a check is unavailable. Safer, but genuine users are stopped too.
Fail open
Letting things through when a check is unavailable, such as approving payments while the fraud scorer is down. Things keep working, but protection is lost.
Failover
Moving work from a failed component to a healthy one, ideally automatically.
Fair scheduling
Sharing a scarce resource between groups by agreed weights, so a busy group can't take everything and a quiet one is never starved.
Faithfulness
How closely an answer sticks to its sources, saying nothing they don't support.
Fallback
A backup path used when the main one fails, such as another model, a cached answer or a simpler feature that doesn't use AI.
False accept
When a face check wrongly says two different people are the same person, letting a stranger through. Teams usually cap it, for example at 1 in 10,000 stranger pairs.
False negative
Something that should have been flagged but slipped through, such as a fraudulent payment that was allowed.
False positive
Something wrongly flagged, such as a genuine payment blocked as fraud.
False reject
When a face check wrongly says a person does not match their own photo, turning away a genuine customer.
Fan-out
Sending one thing to many places, like telling every follower about a new post or every viewer about a new comment.
Fan-out on read
Doing the gathering when someone opens the app: the post is stored once and pulled into each reader's feed only when they look.
Fan-out on write
Doing the copying when something is posted: the new post is added to every follower's ready list straight away, so reading the feed later is quick.
Fare bucket
A group of seats on a flight sold at one price. When the cheapest bucket sells out, the next, dearer bucket becomes the lowest price you can get.
Fault injection
Deliberately causing failures during a test, such as a timeout, an error reply or bad data from a tool, to see whether the system recovers sensibly instead of breaking or repeating an action.
Feature (machine learning)
One piece of information a model uses to make a guess, such as what you bought last week, the time of day or an item's price.
Feature flag
A switch that turns a feature on or off for users without releasing a new version of the app. Teams use it to switch off extras when servers are under pressure.
Feature freshness
How old a feature value is when a model reads it. Teams set a limit per feature, such as "no older than 60 seconds", and alert when it is broken.
Feature registry
A catalogue of every feature with its exact definition, owner, version and freshness target, so teams can find, reuse and trust each other's features.
Feature store
A shared store of the signals a model uses, such as recent clicks or an item's popularity, kept the same for training the model and for answering live requests.
Feed-forward layer
The part of each transformer layer that runs every token's numbers through a large set of learned calculations on their own, without looking at other tokens. It holds most of a model's parameters and much of its stored knowledge.
Feedback loop (in models)
When a model's own decisions shape the data it later learns from. A blocked payment never gets a chargeback, so the next model never learns whether blocking it was right.
Fencing
Cutting off a machine that has been replaced, by blocking its network or taking away its right to write, so it can't keep making changes if it turns out to be still running.
Fencing token
A number that goes up each time a new leader takes over and is attached to every write. Storage refuses writes carrying a lower number than the highest it has seen, so a replaced leader can't change anything.
Few-shot examples
A handful of sample questions with ideal answers placed in a prompt, so the model copies their style and format.
Fill-in-the-middle
A way of training a code model to see the code before and after a gap and write only what belongs in between, which is what grey-text suggestions need.
Filter bubble
When a recommender keeps showing you more of what you already clicked, so you rarely see anything new. It happens because the system mostly learns from what it chose to show.
Fine-tuning
Training an existing AI model further on your own examples so it picks up a habit, such as a fixed format, style or narrow task.
Fingerprint (content hash)
A short code worked out from a file's contents. The same file always gives the same fingerprint, and changing even one pixel gives a different one, so it's a quick way to recognise or check a file.
FinOps
The practice of tracking, sharing out and reducing cloud and AI spending, with engineering, product and finance teams sharing the responsibility.
Firewall
A gatekeeper on a network that allows or blocks traffic by rules, such as letting ordinary web traffic through but stopping most other kinds.
Fixed window
A limit counted per clock period, such as each calendar minute, that resets to zero when the period ends. Simple, but someone can use a full allowance just before the reset and another just after.
FLOPs
Floating-point operations: single multiplications or additions on decimal numbers. Training cost is counted in total FLOPs, and chip speed in FLOPs per second.
Forced update
A screen that blocks an app until you install a newer version. Apps show it when your version is older than the oldest one the server still supports.
Forward secrecy
Using a fresh key for each message or session and deleting old ones, so a key stolen today can't unlock messages sent in the past.
Fraud detection
Automatic checks that look for fake or dishonest activity, like bots playing a video thousands of times, and stop it from counting. They look for patterns real people don't make.
Frequency cap
A limit on how many messages of one kind a person can get in a period, such as no more than 3 offers a day. Messages over the cap are quietly dropped.
Freshness
How old the information behind an answer is. A list built last night has a freshness of up to a day; a live answer is seconds old.
Fulfilment centre
A large warehouse where an online shop stores products and packs orders. Big shops have many, spread across the country, so stock sits close to buyers.
Functional requirement
Something the system must do, such as "people can send a message" or "opening a short link redirects you".
Game day
A planned practice run where a team deliberately causes a big failure, like switching off a whole region, to check that the backup plan really works and that people know what to do.
GDS
Global distribution system: a large shared fare and booking system that many airlines and travel sites connect to, so one site can see many airlines' prices.
Geohash
A short code of letters and digits that names a rectangle on the map. Longer codes mean smaller rectangles, and nearby places usually share the same opening letters.
Geospatial index
A way of organising locations on a map so the system can instantly answer "what is near this point?" without checking every driver in the city.
Goal drift
An agent gradually wandering away from the task it was given, step by step, until it is busy with something the person never asked for.
Gold question
A labelling task whose right answer is already known, mixed in among real tasks so each labeller's accuracy can be measured without them knowing which is which.
Golden set
A fixed collection of test questions with agreed good answers or checks, run on every change to see whether anything that used to work has broken.
Goodput
How many requests a second a system serves within its speed targets, such as a first word within a second. Unlike raw throughput, work that arrives too slowly to be useful doesn't count.
Gossip protocol
A way for machines in a cluster to share news (who's up, who's down) by regularly chatting with a few random peers, like gossip spreading.
GPS ping
A phone telling the server where it is right now, sent every few seconds by a driver's or delivery partner's app.
GPU
A chip with thousands of small cores that does many calculations at once; AI models run on them, and they're expensive.
GPU interconnect
The very fast links between GPUs in the same machine or rack, much quicker than an ordinary network. Splitting every layer of a model across GPUs depends on them.
GPU kernel
A small program that does one piece of maths, such as multiplying two big tables of numbers, on a GPU. How fast a model runs depends a lot on whether good kernels exist for its number format.
GPU memory (VRAM)
The fast memory built into a GPU. A model's weights and its KV cache must fit in it, which often decides how big a model, and how many chats, one GPU can handle.
GPU-hour
One GPU rented for one hour, the usual unit for pricing GPU time. Eight GPUs for a full day is 192 GPU-hours, whether they were busy or not.
Graceful degradation
When a system is overloaded, it switches off the nice-to-have parts and keeps the essential ones working, instead of failing completely.
Gradient
For each parameter, the direction and size of the nudge that would make the model's last answers a little less wrong. Training works these out and applies them over and over.
Gradient-boosted trees
A model made of hundreds of small decision trees, each one correcting the mistakes of those before it. It works very well on tables of numbers and runs fast on ordinary CPUs.
Grading rubric
A written list of what a good answer must do, each with a score scale, so people or a judge model mark AI answers the same way every time.
Graph database
A database built to store things as dots and the relationships between them as lines, so following a chain of links, like a person to their firms to those firms' suppliers, is quick.
GraphQL
A way to build an API where the app describes exactly which fields it wants in one request, and the server returns just those.
GraphRAG
RAG that first builds a knowledge graph from the documents, then answers by following connections or by reading summaries of closely linked groups.
Greedy decoding
Always picking the single most likely next token, with no randomness. It is the same as setting temperature to 0.
Grey failure
A part that is half broken: it still looks healthy to simple checks but is too slow or error-prone for real requests, so it is harder to spot than a crash.
Grievance officer
A named person a large online platform in India must appoint to receive users' complaints, such as a wrongly removed post, and resolve them within deadlines set by law.
Gross margin
The share of revenue left after paying the direct cost of serving customers. AI products usually have thinner margins than ordinary software, because every request costs real money.
Grounding
Tying an AI answer to supplied sources, such as retrieved documents, so it says only what they support.
Group key
A shared key that every member of a group chat receives, each copy locked for that member, so a message can be locked once and read by the whole group.
Grouped-query attention
An attention design where groups of heads share one set of keys and values. It makes the KV cache several times smaller, so a GPU can serve more chats, with little loss in quality.
gRPC
A fast way for programs to call functions on each other over the network, using a shared contract and a compact format.
GSTIN
The 15-character GST registration number printed on Indian tax invoices. It starts with a two-digit state code, includes the business's PAN, and ends with a check character that catches typing and reading mistakes.
Guardrail
A separate check around an AI model that screens what goes in and what comes out, blocking tricks, unsafe replies or private data.
Guardrail metric
A number watched during an experiment to make sure a change doesn't quietly hurt something else, such as complaints, hides or crashes, even when the main number goes up.
Guidance scale
A dial in image generators that sets how strictly the picture follows the prompt. Higher sticks closer to the words but can look harsh; lower looks more natural but may drift from the request.
Hallucination
When an AI model states something confidently that isn't true, because it writes what sounds likely rather than looking anything up.
Handoff
Passing a conversation or task from one agent to another, or to a person, together with the context needed to carry on.
Handshake
The short "hello, hello, OK" exchange two computers do before they start sending data.
Happened-before
The rule that event A came before event B if B could have known about A, for example because A's message reached B's machine first. Events where neither could know about the other are concurrent.
Hash function
A formula that turns any input, such as a user ID, into a scrambled-looking number. The same input always gives the same number, and different inputs spread out evenly.
Hashing
Turning a secret into a scrambled string that can be checked against a guess but can't be turned back into the secret.
Head-of-line blocking
When one stuck item at the front holds up everything behind it. In TCP, one lost packet makes all later packets wait, even ones that have already arrived.
Headless browser
A real web browser that runs on a server with no window on any screen, driven by code instead of a person. It loads pages, runs their scripts and can take screenshots just like a normal browser.
Headroom
Spare capacity you plan on purpose, so a system runs at, say, 60 to 70% busy at its peak instead of 100%. It keeps queues short and absorbs surprises.
Health check
A regular probe a load balancer sends to each server to find out whether it can still take traffic.
Heartbeat
A small "I'm alive" message a machine sends every second or so. When heartbeats stop, others assume it has failed.
Hedged request
Sending the same request to a second copy of a server when the first is slower than usual, and using whichever answer arrives first. It trims the rare slow replies at a small extra cost.
Held-out test set
Examples kept aside and never used for training, so the model can be checked on data it has not seen.
High-level design
The big-picture drawing of a system: its main parts as boxes and the requests between them as arrows, without the fine detail inside each part.
Hint ladder
A fixed sequence of help steps, from a gentle nudge to a fully worked solution, where the app moves a student one rung up only after they have tried the step before.
Hinted handoff
When a machine that should store a write is down, a neighbour keeps the write with a note saying who it belongs to, and hands it over once that machine is back.
Hiring bar
The level of skill a company expects before it hires someone for a role. The same scorecard is used for everyone, but the scores needed rise with seniority.
Hit rate
The share of requests a cache answers from its own copies. At a 90% hit rate, only 1 request in 10 has to go to the slower source.
HNSW
Hierarchical navigable small world: a vector index built as layers of linked points, where a search starts with long jumps on a sparse top layer and zooms in. Fast and accurate, but memory-hungry.
Hold
Keeping something aside for you for a few minutes, like a berth while you pay. If you don't finish in time, it goes back for others.
Holdout group
A small random share of traffic deliberately left out of a system's decisions, so you can see what would have happened without it and get honest data to learn from.
Horizontal scaling
Handling more load by adding more machines (scaling out).
Hot key
One item that everyone wants at once, like a celebrity's post. All the traffic lands on the one place it's stored and can overload it.
Hot path
The route through a system that the most common, most important requests take, such as browsing menus and placing orders in a food app. It is where speed and reliability matter most.
Hotfix
A small, urgent update that fixes one serious bug and changes nothing else, so it can be checked and released quickly.
HTML
The text file that describes a web page: its words, headings and links, and a list of the other files (pictures, styles, scripts) the browser needs to fetch to draw it.
HTTP
The protocol browsers and apps use to send requests and responses over the web, with methods like GET and POST.
HTTP method
The verb at the start of a request that says what you want: GET to read, POST to create, PUT to change, DELETE to remove.
HTTP/2
A newer version of the web's HTTP that lets many requests and replies share one open connection at the same time, instead of opening a new one or waiting in turn.
HTTPS
HTTP sent through a locked, private line, so nobody between you and the website can read or change what's sent. Browsers show it with a padlock next to the address.
Human in the loop
A design where a person must check or approve a step, such as a payment or a deletion, before the software carries it out.
Hybrid logical clock
A timestamp that combines the real clock time with a small counter, so it stays close to real time but never runs backwards or puts an effect before its cause.
HyDE
Hypothetical document embeddings: having a model write a made-up answer to the question and searching with that, because it often looks more like the real answer than the question does.
Hydration
Turning a list of bare IDs into full items by fetching each one's details, such as the caption, author and photo link of every post in a feed.
ICCID
The SIM card's own ID number, usually 19 or 20 digits starting with 89, printed on the card. It identifies the card, not your mobile number.
ID token
A short note signed by the sign-in provider that says who you are, which app it was made for and when it expires. The app checks the signature instead of your password.
Idempotency
Doing something twice has the same effect as doing it once, which makes retries safe.
Idempotency key
A unique label the app attaches to one action, like a payment, and repeats on every retry. The server remembers labels it has seen, so a repeat gets the first answer instead of a second charge.
Identity provider
A service that checks who you are on behalf of other apps, like Google or Apple when you tap "Sign in with Google" or "Sign in with Apple".
IDOR (insecure direct object reference)
A security hole where an app returns a record just because someone asked for its number, without checking it belongs to them. Changing order 1041 to 1042 in a request then shows a stranger's order.
Image preprocessing
The steps that turn an uploaded photo into exactly what a model expects: decode it, turn it the right way up, fix its colours, resize it and turn the pixels into numbers.
Implicit feedback
What users' behaviour suggests without being asked, such as clicking, copying an answer, rephrasing the question or asking for a human. Plentiful but noisy, because the same action can mean different things.
Impossible travel
Two sign-ins to one account from places so far apart, so close together in time, that no one could have travelled between them. At least one of them isn't the owner.
Impression
One time an item, such as a poster or an ad, was actually shown on someone's screen. Counting only items that were really seen keeps the numbers honest.
In-context learning
A model picking up a task from the instructions and examples inside the prompt itself, with no change to its weights.
Incident commander
The one person who leads the response to an incident: deciding priorities, handing out tasks and making sure updates go out, rather than fixing things themselves.
Incident response
A team's planned steps when something goes wrong: spot it, limit the harm, fix it, tell the people who must be told, and learn from it afterwards.
Incremental backup
A backup that sends only what changed since the last one, instead of copying everything again. It is faster and uses less data.
Incremental indexing
Updating a search index with only what is new, changed or deleted, instead of rebuilding it from scratch.
Index
A data structure a database keeps alongside a table so it can find matching rows without reading every row.
Index alias
A stable name that points at one real index. Apps search the alias, so switching to a new index means moving the pointer, not changing code.
Indirect prompt injection
A trick where orders for an AI are hidden inside something it is asked to read, such as an email, a document or a web page, rather than typed by the person chatting.
Inference
Running a trained model to get an answer, as opposed to training it. Every chat reply is an inference call.
Inference server
Software that runs a model on GPUs and answers requests, handling batching, the KV cache and streaming.
Ingest server
The entry point that receives a live video stream from the person broadcasting it, usually in a city close to them. It passes the stream on to be converted and shared out to viewers.
Ingestion pipeline
The background steps that prepare documents for an AI assistant: collect them, pull out the text, clean it, split it into chunks, label them, embed them and store them in a vector index.
Injection attack
Fooling a camera check by feeding a prepared or AI-made video straight into the app, for example through a virtual camera or an emulator, so no real lens is involved.
Inline code completion
Suggestions that appear as you type in a code editor, predicting the rest of the line or function.
Inpainting
Regenerating only a marked patch of a picture and blending it with the rest, such as erasing a stranger and filling in the beach behind them.
Instruction hierarchy
Training that teaches a model to rank instructions by where they come from: the app's own rules first, then the user, and text inside documents or tool results last.
Instruction tuning
Training a pretrained model on examples of requests paired with good answers, so it learns to follow instructions rather than just continue text.
Intent
What a person wants, worked out from what they said, such as 'set a timer for 10 minutes' or 'play songs by this singer'.
Inter-token latency
The gap between one word-piece of a streamed AI answer and the next. Small, steady gaps make a reply feel smooth; long gaps make it stutter.
Interleaving
A way to compare two ranking models by mixing their results into one list for the same person and seeing whose items get picked. It needs far fewer users than a normal A/B test.
Interpolation
Filling in the positions between two known points, so a moving icon glides instead of jumping.
Inventory
The count of things available to sell or book: seats on a train, rooms in a hotel, phones in a sale.
Inverted index
A list of words where each word points to every product or page that contains it, like the index at the back of a book. It is built ahead of time so a search doesn't have to read everything.
IP address
A numeric address that identifies a machine on a network, like 203.0.113.7.
IP location
A rough guess of where you are, usually the city, worked out from your IP address. It can be wrong, especially on mobile data or a VPN.
Issuing bank
The bank that gave you your card or FASTag. It holds the account or wallet behind it and decides yes or no on every payment made with it.
IVF
Inverted file index: a vector index that groups stored embeddings into clusters and, for each question, searches only the few clusters closest to it.
Jailbreak
Clever wording, role-play or framing used to talk an AI model into producing something it has been set up to refuse.
Jitter
A small random extra wait added to retries, so many clients that failed together don't all retry at the same instant.
Jitter buffer
A short wait a call app adds before playing incoming sound and video, so packets that arrive unevenly can be put in order and played smoothly.
Job ID
A reference number given to a piece of slow work the moment it's accepted, like an order slip. The app uses it to check on the work, fetch the result or cancel it.
Join
Combining rows from two or more tables by matching a shared value, such as an order's customer number with the customer's own row, to answer one question.
JSON
A plain-text way to write data that both people and programs can read, like {"item": "dosa", "qty": 2}.
JSON mode
An AI service setting that promises the reply will be valid JSON that any program can read. It doesn't promise which fields the reply contains; that needs a schema.
Judge bias
The ways an AI judge leans unfairly when grading, such as preferring the first answer it reads, longer answers, or answers written by a model like itself.
Just-in-time loading
Keeping only names, ids or links in an AI model's context and fetching the full content with a tool when a step actually needs it, instead of loading everything up front.
JWT
JSON Web Token: a signed pass listing who you are and when it expires, which services can check without a database lookup.
Key agreement
A way for two devices to work out the same secret key from each other's public keys, without ever sending the secret itself across the network.
Key management service (KMS)
A service that creates, stores and controls the encryption keys that lock your data, and logs every use of them. Whoever controls the keys controls who can read the data.
Key pair
Two matching keys made together on a device: a public key that can be shared with anyone, and a private key that stays secret. What one locks or signs, only the other can open or check.
Key salting
Splitting one very busy key into several sub-keys, such as merchant#1 to merchant#16, so its load spreads over many machines, then adding the parts back together.
Key-value store
A simple, very fast database that saves each value under a name, its key, and finds it again by that name, like a locker numbered by product ID.
Keyed state
The running numbers a stream job keeps for each key, such as each phone or each merchant, so every new event only needs to update its own key's numbers.
Keyframe
A still picture that fixes how one moment of a video should look. Video tools often make keyframes first, then animate between or from them.
Kill switch
A control that stops a feature or agent everywhere at once, used when something goes badly wrong.
Knowledge cutoff
The point where a model's training text ends. The model knows nothing that happened later unless the app supplies it, for example through search or RAG.
Knowledge graph
A store of facts as things, called entities, joined by named relationships, such as 'Asha works at Acme'.
Knowledge tracing
Keeping a running estimate, per student and per topic, of how likely the student is to have mastered it, updated after every answer they get right or wrong.
KV cache
The keys and values a model has already worked out for every earlier token, kept in GPU memory so each new token doesn't mean re-reading the whole text. It grows with context length and with the number of chats served at once.
KV offloading
Moving an idle chat's KV cache out of scarce GPU memory into the server's ordinary memory or disk, and copying it back when the chat resumes. Copying back is usually much faster than recomputing the notes.
KYC
Know your customer: the identity check a bank must do before it opens an account or wallet for you, using documents like Aadhaar or PAN.
Label bias
A model's tilt towards a label because it appeared more often, or last, in the examples it was shown, rather than because the new input deserves it.
Label delay
The gap between a model's decision and learning whether it was right. In payments, a stolen card may be reported weeks later, so the model can only learn from older payments.
Label noise
Wrong answers among a set of training labels, from mistakes, unclear rules or careless work. A little is normal; a lot makes the model learn the wrong thing.
Labeller agreement
How often two people labelling the same examples give the same answer. Low agreement usually means the labelling guidelines are unclear, so the labels can't be trusted yet.
Labelling guidelines
The written rules labellers follow, with example items on each side of every tricky line, so that hundreds of people mark the same item the same way.
Lamport clock
A simple logical clock: each machine counts its events, stamps messages with its count, and on receiving one jumps past the bigger number. Causes always get smaller numbers than their effects.
LangChain
An open-source framework of building blocks for AI apps, such as prompt templates, model connectors, retrievers and tools, that can be chained together.
LangGraph
An open-source framework for building agents as graphs of steps with saved state, so loops, branches and pauses for people are spelled out.
Last write wins
A simple rule for two changes to the same thing at nearly the same time: the change that arrives last is the one that's kept.
Late event
An event that reaches the system well after it happened, for example from a phone that was offline, so it arrives after newer events and after its time window seemed finished.
Late interaction
A search method that keeps one embedding per token and matches question and passage token by token at search time, catching exact details that a single embedding blurs.
Latency
How long one request takes, from sending it to getting the answer back. Usually measured in milliseconds (ms).
Latency budget
The total time a system is allowed to take, split into a share for each step, like a household budget for time. If one step overspends, another must save.
Latent space
A compressed form where an image or sound is described by a much smaller set of numbers. Image models do most of their work there and turn the result into pixels at the end.
Layer 4 load balancer
A load balancer that spreads network connections without reading what's inside them. Very fast, but it can't route by web address or see individual requests.
Layer 7 load balancer
A load balancer that reads each web request, so it can send it to different servers by its address, headers or cookies, and balance one request at a time.
Layout analysis
Working out a page's structure, such as headings, columns, tables and form fields, so its text can be read in the right order and context.
Leader
The one machine in a group that is allowed to accept changes; the others follow and copy it.
Leader lease
A time-limited right to act as leader, renewed every few seconds. A leader that can't renew it must stop acting before it runs out, and the others wait for it to expire before choosing a new one.
Leaky bucket
A way to pace work: requests wait in a queue that drains at a fixed rate, so bursts are smoothed out into a steady flow instead of being refused.
Learning rate
How big each adjustment to a model's numbers is during training. Too high and the model forgets what it knew or starts babbling; too low and training barely changes anything.
Lease (on a job)
A time-limited claim a worker takes on a job while working on it, renewed as it goes. If the worker dies and the lease runs out, the job goes back in the line for another worker.
Least connections
A way for a load balancer to choose a server: send each new request to the one with the fewest requests in progress, so busy or slow servers get less new work.
Least privilege
Giving each person, program or tool only the access it needs for its job and nothing more, so a mistake or attack can do less harm.
Ledger
An append-only record of every money movement, written as balanced debits and credits, never edited, only corrected with new entries.
Lethal trifecta
The dangerous mix of an AI agent that can see private data, reads untrusted content and can send data out. With all three, a hidden instruction can leak your data.
Letterboxing
Fitting a photo into a square by shrinking it whole and filling the empty sides with a plain border, instead of stretching or cropping it.
Lifecycle rule
A rule in file storage that acts on files by age automatically, such as deleting drafts after 30 days or moving old files to cheaper storage.
Lift
How many times more often two items are bought together than chance alone would predict. A lift of 1 means no real link; a lift of 50 means a strong one, even if the raw count is small.
Linked device
A laptop, tablet or second phone connected to your chat account. Each one has its own keys and gets its own copy of every message.
Lip-sync
Adjusting a speaker's mouth movements in a video so they match a voice track, for example after dubbing into another language.
Little's law
A simple rule for any line of work: items arriving per second times seconds each item takes equals the number in progress at once. 4 jobs a second at 6 seconds each means about 24 being worked on.
Liveness check
A health check that asks whether a server is stuck. If it fails, the server is restarted.
liveness detection
A check that the face or finger in front of a camera or scanner belongs to a living person who is there right now, not a photo, a mask or a rubber copy.
LlamaIndex
An open-source framework focused on connecting AI models to your data: loading documents, indexing them and retrieving from them.
LLM
Large language model: an AI model trained on huge amounts of text that writes answers one word piece at a time.
LLM judge
A second AI model asked to grade another model's answers against a written rubric. Its grades are checked against people's grades before they're trusted.
Load balancer
A component that receives requests and spreads them across several servers.
Load shedding
Deliberately turning away some requests when a system is overloaded, so it can keep serving the rest well.
Load test
A rehearsal where a system is flooded with fake users on purpose, to find out how much it can take and what breaks first, before real users arrive.
Local storage
Space on your phone or in your browser where an app or website can save small things, like a guest cart, so they're still there after you close it.
Lock
A temporary "reserved" tag on something, like a seat, so only one person can change it at a time. Everyone else waits or is told it's taken.
Lockfile
A file that records the exact version of every library an app uses, including the libraries those depend on, so every install gets the same code until someone deliberately upgrades.
Log compaction
A setting that keeps only the latest event for each key and drops older ones, so the log works as an always-current list, like the latest price of every product.
Log probabilities (logprobs)
The model's own probability for each token it wrote, reported in a mathematical form. Low values point to places where the model was unsure.
Log retention
How long recorded logs and traces are kept before they're deleted. Shorter periods for full conversations reduce the harm if stored data ever leaks.
Logical clock
A counter that goes up with every event, used to tell what happened before what without trusting real clocks.
Long tail
The huge number of items, like old videos or rare products, that each get very few views or sales. Together they add up to a lot, but no single one is popular enough to keep everywhere.
Long-term memory (AI)
Facts an AI app saves about a person across chats, such as their name or that they're vegetarian, and looks up again when a new question is related.
Look-to-book ratio
How many searches a travel site handles for every booking it makes. It is often hundreds to one, which is why sites save and reuse search results.
Lookalike domain
A web or email address made to look like a real one at a glance, such as rnybank.in for mybank.in, used to trick people into trusting a fake message.
Loop detection
Code around an AI agent that notices when it repeats the same step with the same inputs, and interrupts it so it tries another way or stops.
Loop engineering
Designing, measuring and improving the harness around an agent, rather than only its prompt or model.
LoRA
Low-rank adaptation: training a small set of add-on weights beside a frozen model. The add-on is tiny compared with the model and can be swapped in and out.
Loss
A single number that measures how wrong a model's guesses were on a batch of training examples. Training nudges the model's numbers so that this number goes down.
Loss spike
A sudden jump in the training loss. It can come from bad data, numbers overflowing or a setting that is too aggressive, and sometimes the run never recovers.
Lost in the middle
The tendency of models to use facts at the start and end of a long context well but miss ones buried in the middle.
LRU eviction
Least recently used: when a cache is full, it removes the item that was used longest ago to make room for a new one.
LSM tree
A way of storing data built for fast writes: new writes go into a sorted table in memory, which is written to disk in one go when full, and background work keeps merging the files on disk.
Machine learning
Software that learns patterns from a huge number of examples, instead of following only rules a person wrote down. Shown millions of scam emails, it learns what a scam tends to look like.
Machine translation
Software, usually an AI model, that turns text in one language into another, such as English commentary into Tamil.
Man-in-the-middle attack
An attacker secretly sitting between two people, posing as each to the other, so they can read or change messages that both sides think are private.
Manifest
A small file listing every piece of a video and the qualities each one comes in. The player reads it first.
Many-shot prompting
Putting hundreds of worked examples into a prompt instead of a handful. With a large window and prompt caching it can rival fine-tuning, and changing an example is just a text edit.
Map cell
The map split into small tiles (often hexagons) so things that are close together share a tile. Searching nearby means checking your tile and its neighbours.
Map matching
Snapping rough GPS points onto the road a vehicle is actually on.
Map tile
A small square piece of a map, drawn once and saved, so phones can fetch the ready-made pieces they need instead of having the whole map drawn for each of them.
Map-reduce
Splitting a big job into pieces handled separately (map) and then combining the results (reduce), such as summarising each chapter and then summarising the summaries.
Margin of error
How far a test score could easily move just from luck in which questions happened to be picked. On a few hundred questions it is often several points either way, so smaller gaps between models may mean nothing.
Markdown
A simple way of marking formatting in plain text, such as **bold** or a dash for a bullet point, which apps turn into styled text.
Masking
Hiding most of a personal value while keeping a small hint, such as showing a card as XXXX XXXX XXXX 1234, so a person can recognise it but nobody can use it.
Mastra
An open-source TypeScript framework for AI agents and workflows, with typed tools, memory, evals and a local playground for testing.
Matching
Pairing two sides, like each rider with the best free driver nearby.
Materialisation
A scheduled job that computes feature values and copies the latest ones into the online store, so live requests can read them instead of calculating them.
Maths solver (computer algebra system)
Ordinary software that does algebra and arithmetic exactly, by rules rather than by guessing, so it can check whether a step like 2x + 5 = 17 to 2x = 12 is correct.
MaxSim
The scoring rule used in late-interaction search: for each word of the question, take its best match among the passage's words, then add up those best scores.
MCP
Model Context Protocol: an open standard that lets any app offer tools, data and prompt templates to any AI app through one common kind of connection.
MCP client
The AI app side of an MCP connection, such as a chat app or coding tool. It asks MCP servers what they offer, shows their tools to the model and passes the model's tool calls on.
MCP server
A program that offers tools or data to AI apps over MCP, such as one for a company's calendar or tickets. It can run on your own machine or remotely.
Mean reciprocal rank
A search score for how high the first right result appears: 1 if it is first, a half if second, a third if third, and 0 if it is missing, averaged over many test questions.
Media fingerprint
A summary of what a video or song looks and sounds like, used to spot copies even when they're cropped, re-recorded or slightly changed. Unlike an exact file fingerprint, small edits don't fool it.
Memory (RAM)
A computer's short-term working space. Reading from it takes about a ten-millionth of a second, far faster than any disk, but it is wiped when the power goes off.
Memory bandwidth
How quickly a chip can move data out of its memory. Writing each token means reading the model's weights, so bandwidth often limits decode speed more than raw calculation does.
Memory consolidation
A background job that tidies an AI assistant's saved memories: merging duplicates, replacing facts that have changed and retiring stale ones, so the memory stays accurate as it grows.
Memory fragmentation
Free memory broken into scattered gaps, each too small to use, so there is room in total but not in one piece. Reserving more than is needed makes it worse.
Memory poisoning
An attack that slips false or malicious entries into an AI assistant's long-term memory so they sway later conversations.
Memory table (memtable)
The sorted table in memory where an LSM-tree database collects the newest writes until it is full enough to be saved to disk as one file.
Merkle tree
A tree of fingerprints where each parent summarises the fingerprints below it. Two machines can compare huge sets of data by comparing the top fingerprint first and digging only into branches that differ.
Message passing
Parts of a system working together by sending each other messages, so each part knows only what it has been sent, rather than all reading and writing one shared store.
Message queue
A buffer that holds messages until a consumer is ready to process them, so producers don't have to wait.
Message roles
The labels on each part of a chat sent to an AI model: system for the app's rules, user for what people typed, assistant for the model's earlier replies. Models give each role a different level of trust.
Message template
A saved message with blanks, like "Your order from {shop} is on the way", kept in each language. The details are filled in just before sending.
Metadata
Small facts about a bigger thing: a photo's owner, caption, size and where it's stored, but not the photo itself.
Metadata filter
A rule added to a search that only lets through items whose labels match, such as "tenant is Company A". Items that fail it are never returned.
Metric
A number recorded again and again over time, such as requests a second or the share of failed payments, cheap to store and easy to chart or alert on.
Micro VM
A very small virtual computer that starts in a fraction of a second and is walled off from other ones at the hardware level, so untrusted work can't reach its neighbours.
Microservice
A small, separate program that does one job (payments, search, delivery) and talks to the others over the network.
Mixed precision
Training with 16-bit numbers for most of the maths, to save memory and time, while keeping a 32-bit master copy of the weights so small updates are not lost to rounding.
Mixture of experts (MoE)
A model design with many specialist sub-networks, called experts, where a small router picks only a few of them for each token. A very large model then costs far less to run per token.
Mock tool
A pretend version of a real tool, such as a payments or booking service, used in tests. It gives realistic, controlled answers, and can fail on purpose, without touching real money or real people.
Model alias
A named pointer in a model registry, such as "production", that points at one model version. Moving the pointer promotes or rolls back a model without rebuilding anything.
Model artifact
The saved file (or set of files) that is the trained model itself, such as its learned weights or trees. Servers load it to make predictions.
Model card
A short document published with a model that says what it is for, how it was trained and tested, its known limits and where it shouldn't be used.
Model cascade
Trying a cheap model first and passing the request to a stronger, dearer one only when the first is unsure or fails a check.
Model collapse
What happens when models are trained, generation after generation, on text written by models: their output grows narrower and blander and their mistakes pile up. Mixing in real human data and checking answers guards against it.
Model extraction
Copying an AI model's skills by sending it huge numbers of questions and training another model on its answers, usually against the provider's terms.
Model FLOPs utilisation (MFU)
The share of a GPU's top advertised speed that a training run actually turns into useful model maths. Large runs commonly reach about 30 to 50%.
Model registry
A catalogue of every trained model version with the data, code and test results that produced it, and which version is live, so you can compare, promote or roll back in one step.
Model routing
Choosing which AI model handles each request, for example sending easy questions to a small, cheap model and hard ones to a stronger one.
Model stage
A label in a model registry that says where a model version is in its life: a candidate, in staging for testing, in production, or archived.
Model version pinning
Naming an exact model version in your app, rather than a name the provider may point at a newer model, so behaviour doesn't change without warning.
Monolith
An app built and deployed as one single program, rather than as many small services. It's simpler to build and run, but every part must scale and be released together.
Monotonic reads
A promise that once you have seen a change, you won't later see an older version, usually kept by sending each user to the same copy of the data.
MTTR (mean time to recover)
The average time from a problem starting to service being back to normal. Faster detection and easy rollback shrink it more than anything else.
Mule account
A bank account, often rented or opened with someone else's papers, that fraudsters use to receive stolen money and pass it on quickly.
Multi-adapter serving
Keeping one base model loaded on a GPU and applying a different adapter per request, so many fine-tuned variants share the same hardware.
Multi-agent system
A design where several AI agents, each with its own role, tools or context, work together on one task.
Multi-armed bandit
A way of choosing between several options while still learning which is best: mostly pick the option that has worked so far, and sometimes try the others to keep learning.
Multi-hop question
A question that needs several lookups in a row, where the answer to one tells you what to search for next.
Multi-label classification
A model that gives a separate yes-or-no score for each label, so one item can be several things at once, such as both blurry and carrying a phone number.
Multi-tenancy
One app serving many customer companies, where each company's data must stay strictly separate.
Multimodal embedding
An embedding that places images and text in the same space, so a sentence can find a matching picture and the other way round.
Multimodal model
An AI model that can take in or produce more than text, such as pictures, voice or video.
Multimodal RAG
RAG that can search and answer from more than plain text, such as tables, charts, photos and whole page images, while still pointing back to the page each answer came from.
N+1
Running one more copy than you need at your busiest time, so the rest can still carry the load when one copy fails.
N+1 query problem
Running one query to fetch a list and then one more query for each item in it, so a page of 50 items makes 51 trips to the database instead of one or two.
Near-duplicate
Two pieces of text that are almost the same, such as one FAQ copied onto two pages with only the date changed. Exact matching misses them, so they need a fuzzier comparison.
Near-real-time
Results that are updated within seconds or minutes of something happening, rather than instantly on each request or once a night.
Network partition
When the network between parts of a system breaks, so some machines can't reach others even though each is still running.
Neural network
A program built from many layers of simple sums whose numbers are adjusted during training until it gets good at a task. Modern AI models are very large neural networks.
Next-token prediction
The one job a language model is trained to do: given the text so far, give every possible next token a probability. A whole answer is this step repeated.
Noisy neighbour
One customer on a shared system using so much of it that everyone else slows down.
Non-functional requirement
How well the system must do its job: how fast, for how many people, how reliably, and how fresh the data must be.
NoSQL database
A family of non-relational databases (key-value, document, wide-column, graph) that trade some features for flexibility or scale.
Novelty effect
People reacting to something just because it is new, such as clicking more on a changed home screen in its first week. The lift fades once the change stops feeling new.
NPU
Neural processing unit: a chip in phones and laptops designed to run AI models quickly on little power.
NTP (Network Time Protocol)
The standard way computers set their clocks by regularly asking time servers. It keeps machines close to real time, usually within milliseconds, but never perfectly in step.
Number portability
Moving your mobile number from one operator to another without changing it. A central record of which operator serves each number is updated when you switch.
number recycling
Giving a mobile number that someone gave up to a new customer, after it has sat unused for a cooling-off period of a few months.
OAuth
The standard behind "Sign in with Google": you sign in on Google's page and the app receives a token, never your password.
Object detection
An AI model that finds the separate things in a photo, such as a shoe, a bag and a phone, and draws a box around each one with a label and a score.
Object storage
A cheap, practically unlimited store for big files like photos and videos. You save and fetch each file by a key, like a cloakroom ticket.
Observability
How well you can see what's happening inside a running system, through metrics, logs and traces.
OCR
Optical character recognition: software that reads the letters in a photo or scanned page and turns them into text a computer can search.
Off-policy evaluation
Estimating how a new decision rule would have performed using logs collected under the old rule, by reweighting each logged choice by how likely it was. It is a cheap check before a live A/B test.
Offline evaluation
Scoring a model on saved, past data before it meets real users. It shows whether the model's answers are accurate, but not how people will react to them.
Offline pack
A downloadable bundle for one chapter, with explanations, worked examples, practice questions and their checks, so the app can keep teaching when there is no network.
Offline store
The part of a feature store that keeps the full history of every feature value with its time, in a data warehouse, so models can be trained on past examples.
Offline-first
Building an app so it works from what's saved on the phone first and catches up with the server later. No internet is treated as normal, not as an error.
Offset
An event's position number inside one partition of an event log. A reader remembers the offset it has reached, like a bookmark, so it can carry on from there after a restart.
Offset
An event's position number inside one partition of an event log. A reader saves the offset it has reached so it can carry on from there after a restart.
On-call
Taking turns being the engineer who is reachable at any hour to respond when an alert fires, usually for a week at a time.
On-demand instance
A cloud server rented by the hour or second with no commitment, which you can stop at any time. It is the most flexible way to rent and usually the dearest per hour.
On-device AI
Running an AI model on your own phone or laptop instead of on a company's servers, so your data doesn't have to leave the device.
On-time rate
The share of orders that arrive by the time the app promised. Delivery apps watch it more closely than average error, because a late order hurts far more than an early one.
Online store
The part of a feature store that keeps only the latest value of each feature in a fast key-value database, so a live request can fetch it in a few milliseconds.
Open-weight model
An AI model whose trained files are published, so anyone can download it and run it on their own servers instead of using a provider's service.
OpenID Connect
A layer on top of OAuth that lets an app learn who you are from a sign-in provider like Google, as a signed ID token with your name and email.
OpenTelemetry
An open standard, with free libraries, for recording traces, metrics and logs in a common format, so an app's data can be sent to any monitoring tool without rewriting the app.
Operating point
The threshold a team actually runs a model at, and the mix of catches and false alarms that comes with it. It is chosen from what each kind of mistake costs.
Operation
One small, precise change to a document, like "insert the letter a at position 120" or "delete 3 characters at position 88". Shared editors send these instead of the whole document.
Operational transformation (OT)
A way to merge changes from several people editing at once: a change made on an older version is adjusted (for example, shifted along by a few positions) so it still fits after the changes that came before it.
Optimiser state
Extra numbers the training algorithm keeps for every parameter, such as a running average of recent gradients. With the popular Adam optimiser it takes more memory than the weights themselves.
Optimistic concurrency
A way to stop two changes clashing without locking: each change says which version it expects to replace, and it is refused if someone else changed the data first, so it must reload and try again.
Optimistic locking
A way to avoid two people overwriting each other without locking anything: each record carries a version number, and a change is saved only if the version is still the one you read. If not, you reread and try again.
Optimistic update
Showing the result on screen straight away (the heart turns red) and saving it in the background, fixing it only in the rare case saving fails.
Orchestrator
The agent or program in a multi-agent system that splits up the work, hands pieces to other agents and combines their results.
Origin server
The main server that holds the original copy of a video or file. CDN servers fetch from it once and then serve copies to everyone nearby.
origin shield
A middle layer of large caches between the CDN's many edge servers and the origin. Edges ask the shield, and only the shield asks the origin, so the origin sees a handful of requests instead of thousands.
OTP
One-time password: a short code, usually 6 digits, sent to your phone. It works once, and only for a few minutes.
Outbox pattern
Saving "send this message" in the same database transaction as the change itself, then sending it from there, so it can't be forgotten.
Outcome-based pricing
Charging for a finished result, such as each support ticket resolved, rather than for seats or usage. It only works if you know what each result costs you.
Outlier
A reading that doesn't fit the rest, like a parked car among moving ones. Systems often drop outliers so they don't spoil an average.
Output limit
The most tokens an app lets an AI model write in one answer. When the limit is reached, the answer stops, even in the middle of a sentence.
Overfitting
When a model learns its training examples too closely, quirks included, and does worse on new inputs. A held-out test set reveals it.
Overselling
Selling more of something than you actually have, like confirming 1,200 orders for 1,000 phones. Booking and sale systems are built to make this impossible.
OWASP
The Open Worldwide Application Security Project, a non-profit security community known for its Top 10 lists of the most serious risks, including one for apps built on language models.
p50 latency (median)
The middle response time: half of requests are faster than this and half are slower. It shows the typical experience but hides the slow ones.
p99 latency
The 99th percentile: 99 out of 100 requests are at least this fast. It describes the slowest experiences users actually have.
PACELC
An extension of the CAP trade-off: during a network split you choose between availability and consistency, and Else, in normal times, between speed (latency) and consistency.
Packet
A small piece of data travelling over the internet. Anything bigger, like a photo, is cut into many packets.
Page (in a database)
The fixed-size block, usually 4 to 16 KB, that a database reads from and writes to disk in one go. Changing one small row means rewriting its whole page.
Page split
What a B-tree does when a new entry must go into a page that is already full: it splits the page into two half-full pages and updates the page above to point to both.
Page table
A list kept for each chat that records which small memory pages hold its notes, in order. The pages can sit anywhere in memory; the table keeps track of them.
Paged attention
Storing the KV cache in small fixed-size pages, the way an operating system manages memory, so little memory is wasted and more chats fit on a GPU.
Parameter-efficient fine-tuning (PEFT)
Fine-tuning methods that train only a small number of new or chosen parameters and leave the rest frozen, cutting the memory and storage needed.
Parameters (weights)
The billions of numbers inside an AI model that were tuned during training; together they hold what the model has learned. A model's size is usually given as its number of parameters.
Parent–child chunking
Searching small chunks, which match a question sharply, but handing the AI model the larger section each one came from, so it gets enough context to answer well.
Partial result
An early guess from streaming speech to text while someone is still talking. It can change a moment later, when the model hears more and gives its final version.
Partition
One slice of an event stream or table, so work can be spread across machines; order is kept within each slice.
Pass rate
The share of repeated runs in which an agent completes a task. Because agents vary from run to run, each test is run several times rather than once.
pass@k and pass^k
Two ways to score an AI system that is run k times on the same task. pass@k counts a success if at least one of the k tries works; pass^k counts it only if all k tries work, which is closer to how reliable it feels to a user.
Passkey
A way to sign in without a password: your phone or laptop keeps a secret key and proves it's you by signing a challenge, unlocked with your fingerprint, face or screen lock. It only works on the real site, so fake pages can't steal it.
Password manager
An app or browser feature that creates a different strong password for every site and remembers them all for you, so you only need to remember one.
Payment gateway
The service that talks to banks, cards and UPI apps to actually move your money, then tells the app whether it worked.
Peak factor
How much busier the busiest hour is than the daily average. A peak factor of 3 means you must plan for three times the average load.
Peeking
Checking an experiment's result again and again while it runs and stopping as soon as it looks like a win. It makes lucky flukes look like real wins far more often.
Pending
A payment whose final result isn't known yet. It will settle as success or failure; the system keeps checking until it does.
Perceptual hash
A short fingerprint of what a picture looks like, so resized, cropped or slightly edited copies still match a known image. Used to spot known illegal images without anyone having to look at them again.
Permission mode
A setting that decides what an AI agent may do on its own and what it must ask a person about first, such as editing files or running commands.
Permission-aware retrieval
Filtering search results by what the person asking is allowed to see, at the moment of searching, so an AI assistant never reads documents they couldn't open themselves.
Persistent connection
A line between your phone and the server that stays open, so either side can send something the moment it happens instead of asking again and again.
Personal data (PII)
Any information that identifies a person or could be linked to them, such as a name, phone number, address or Aadhaar number.
Phishing
A fake website, message or call made to look like a real one, to trick you into typing a password, OTP or card details.
Pick list
The list of items a store worker must collect for an order, usually sorted in the order of the shelves so they can walk the store just once.
Pinball loss
A training score that punishes guesses that are too low more than guesses that are too high, or the other way round, by a chosen ratio. Training with it teaches a model to predict a percentile.
Pipeline bubble
Time a GPU in a pipeline sits idle at the start and end of each step, waiting for work to reach it or come back to it. More, smaller batches shrink it.
Pipeline parallelism
Splitting a model's layers into groups, each on its own GPU, with work passing from one to the next like an assembly line.
Pivot step
The point of no return in a Saga, often charging the card. Steps before it are easy to undo; steps after it are retried until they succeed instead of being undone.
PKCE
A safety check for phone and browser apps that can't hide a password: the app makes a fresh random value for each sign-in and must show it to swap the one-time code, so a stolen code is useless.
Plain text
Stored exactly as typed, with no scrambling, so anyone who sees the stored copy can read it.
Plan and execute
An agent design that writes a full plan first, then carries out the steps one by one, re-planning only when something changes or goes wrong.
PNR
Passenger Name Record: the 10-digit number on a train ticket that identifies one booking and everyone travelling on it.
Point-in-time correctness
Building each training example only from feature values that were known at the moment of that example, never from values recorded later.
Poison message
A job or message that fails every time it's handled, for example because its data is broken. Retrying it never helps, so it has to be set aside.
Policy engine
A separate service that holds an organisation's access rules and answers 'may this person do this action on this item?' so every app checks the same rules the same way.
Polling
Asking the server again and again, every few seconds, whether anything has changed.
Popularity bias
When items that are already popular get shown more, so they collect more clicks and get shown even more, while new or niche items never get a fair chance.
Population stability index (PSI)
A single number that measures how much a distribution, such as payment amounts or model scores, has shifted between a reference period and now. Bigger means more change.
Position bias
People tap items near the top of a screen more often simply because they see them first, so a high spot can look like popularity when it isn't.
Position encoding
Extra information added inside a transformer so each token knows where it sits in the text. Without it, the model couldn't tell 'dog bites man' from 'man bites dog'.
Post-training quantisation
Shrinking a finished model's numbers to fewer bits after training, using a small sample of typical inputs, without retraining it. Quick and cheap, but it loses a little more quality than planning for the rounding during training.
Postmortem
A written review after something goes wrong: what happened, when, why, and what will change so it doesn't happen again. Good ones look for causes in the system, not people to blame.
Pre-labelling
Having a model fill in its guess for each item first, so labellers only confirm or correct it instead of starting from scratch.
Pre-warming
Getting servers, caches and pages ready before a known rush, so the system isn't cold and slow when the crowd arrives.
Precision
Of the items a system picked out, the share that were actually right, such as how many flagged posts really broke the rules.
Precision (bits per number)
How many bits are used to store each number in a model, such as 32, 16, 8 or 4. Fewer bits mean less memory but rougher numbers.
Precomputation
Working out answers ahead of time and saving them, so that when someone asks, the system just looks the answer up instead of calculating it.
Prediction log
A record of every answer or decision a model made, with the input, the context it saw, the model and prompt version, and a unique ID, so later feedback can be tied back to it.
Prediction store
A fast key-value database that holds model answers worked out ahead of time, such as each user's recommended list, so the app can read them in a millisecond instead of running the model.
Preemption
Pausing a running request to free its resources for others, then resuming it later. AI servers do this when GPU memory runs out, either saving the request's notes elsewhere or rebuilding them on resume.
Preference tuning
Training a model on which of two answers people (or a judge) preferred, so it leans towards helpful, safe and well-judged replies.
Prefill
The first phase of answering, where the model reads the whole prompt in one go, in parallel. Longer prompts make prefill take longer, which delays the first word of the reply.
Prefix caching
A serving-engine technique that keeps the KV cache for a shared prompt opening and reuses it for every request that starts the same way. It is the mechanism behind prompt caching.
Prekey
A spare public key a phone leaves with the chat server in advance, so someone can start a locked conversation with it even while the phone is switched off. Single-use ones are handed out only once.
Presence
Knowing whether a user is online right now, and which server their phone is connected to. It is kept in a fast shared store and refreshed by regular heartbeats, so it can be slightly out of date.
Presentation attack
Fooling a camera check by holding something up to the lens, such as a printed photo, a video playing on another phone or a mask.
Presentation timestamp
A time stamp carried inside the audio and video that says exactly when each frame or sound should play, so separate tracks like captions can line up with it.
Pretraining
The first and biggest training stage, where a model learns to predict the next token across a huge collection of text. It takes vast compute and gives the model its general knowledge.
Prevalence
The share of everything people actually saw that broke the rules, measured by having people label a random sample of views. It shows how much harm got through, not just how much was removed.
Price quote
A fare worked out for one exact trip and kept valid for a few minutes, so the price you agree to is the price you pay.
Primary database
The main copy of a database. Every change is saved here first, then copied to the replicas.
Primary key
A value that is unique to each row of a table, such as an order number, so the database can always find exactly one row by it.
Primary metric
The one number an experiment is designed to move, agreed before it starts, such as orders per user or late deliveries. The decision is made on this number.
Priority queue
A waiting line with more than one lane, where jobs in the higher lane are picked first. Often used to give paying customers shorter waits.
Private key
The secret half of a key pair, kept only on your own device. It is the only key that opens messages locked with your public key.
Privilege separation
Splitting work so the part that reads untrusted content has no power to act, and the part that can act never sees the raw untrusted text.
Procedural memory
Remembered know-how about doing tasks, such as a person's preferred steps for filing expenses, saved as instructions.
Processing time
The moment a system reads and handles an event, by its own clock. Simple to use, but wrong whenever events arrive late or are replayed from history.
Producer
Any service that writes events into an event log or queue, like an order service writing "order placed".
Product quantisation
A way to shrink stored embeddings: each one is cut into short pieces and every piece is replaced by a tiny code for its closest stock pattern. Memory drops many times over, at the cost of slightly blurrier distances.
Project instructions file
A plain text file kept in a code repository that tells AI coding tools how the project works: how to run its tests, its rules and what not to touch.
Projection
A small program that reads a stream of events in order and keeps a ready-made view up to date, such as a balance or a monthly statement. The view can always be rebuilt from the events.
Prompt
The full text an app sends to an AI model on one call: instructions, examples, facts, recent chat and the user's message.
Prompt bombing
Sending someone sign-in approval prompts again and again, hoping they tap Yes by mistake or just to make them stop.
Prompt caching
A model provider remembering its work on the opening part of a prompt, so later requests that start with exactly the same text are cheaper and faster.
Prompt chaining
Splitting a job into a chain of model calls where each one's output feeds the next, such as outline, then draft, then polish.
Prompt injection
Text written to trick an AI model into ignoring its instructions, typed into a chat or hidden in a document or web page the model reads.
Prompt lookup decoding
A way of guessing a model's next few tokens by finding the words it just wrote in the prompt and proposing whatever followed them there. It needs no extra model and works best for code edits and quoting.
Prompt registry
A central store that keeps every version of every prompt, with its model and settings, who changed it and which version is live for which users. Apps ask it which prompt to use instead of having the text written into their code.
Prompt template
A prompt with blanks, such as the customer's name or order details, that the app fills in on every request.
prompt worm
A prompt injection written to copy itself into whatever an AI assistant writes next, such as an email, a file or a memory note, so it spreads to the next assistant that reads it.
Propensity
The probability a system had of making the choice it made, logged at decision time. It lets you later reweight old logs to judge how a different choice rule would have done.
Protocol Buffers
A compact binary format for sending data between programs. Both sides share a contract file that gives every field a number, so messages carry numbers instead of spelled-out field names.
Proxy
A server that sits between you and another server and passes requests and answers along, sometimes changing, holding or checking them on the way.
Proxy metric
Something you can measure easily that stands in for what you really care about, such as clicks standing in for whether people liked a video. It can mislead if optimised on its own.
Public key
The shareable half of a key pair. It can check signatures made with the matching private key, which stays secret, but can't make new ones.
Publish–subscribe
Senders publish messages to a named channel, and everyone subscribed to that channel receives them. The sender doesn't need to know who is listening.
Pull request
A proposed set of code changes sent for review before they are merged into the main project.
Purpose limitation
The rule that personal data collected for one stated purpose, such as processing a loan, is used only for that purpose and not quietly reused for something else.
Push notification
A message the app's server sends to your phone even when the app is closed, like "Your order is on the way".
Push service
Google's service on Android phones, or Apple's on iPhones, that keeps one connection open to each phone and delivers notifications from every app along it.
Pydantic AI
An open-source Python framework for building AI agents as ordinary typed code, where the model's reply must match a declared shape and tools are defined from normal functions.
QLoRA
LoRA trained on top of a quantised, compressed copy of the base model, so fine-tuning a fairly large model fits on a single GPU.
QR code
A square pattern of black and white blocks that stores a short piece of text, like a ticket number or a payment address, for a camera to read.
Qualified play
A play that counts as real interest, such as watching at least 10 minutes of a film or most of an episode, rather than a tap followed by a quick exit.
Quantile prediction
Predicting a percentile instead of the average, such as the time that 85 in 100 similar orders beat. It lets a promise lean towards the safe side by a chosen amount.
Quantisation
Storing a model's numbers with fewer bits, such as 4 instead of 16, so it uses less memory and runs faster, at the cost of some accuracy.
Quantisation-aware training
Training a model with the later rounding to fewer bits switched on, so it learns to cope with it. Such models lose less quality when shrunk.
Query plan
The database's chosen way to answer a query, such as which index to use or whether to read the whole table. Most databases will show it on request, which is how slow queries are diagnosed.
Query rewriting
Turning a user's question into a clearer search, for example by adding missing context from the chat, before looking anything up.
Query routing
Choosing which source or index to search for a question, such as the policy documents rather than the order database.
Query, key and value
The three lists of numbers each token makes inside attention: a query for what it is looking for, a key for what it offers, and a value for the information it passes on when picked.
Queue depth
How many jobs are waiting in a queue right now. When it keeps growing, workers can't keep up, which is the usual signal to add more of them.
QUIC
A newer way of sending data, built on UDP, that adds its own reliability and encryption. It connects in fewer round trips, keeps one lost packet from holding up everything, and is what HTTP/3 runs on.
Quorum
The minimum number of copies (usually a majority) that must agree before a read or write counts.
Quota
A share of seats set aside for a particular group or purpose, such as Tatkal, ladies, senior citizens or passengers from certain stations, with its own booking rules.
RAC
Reservation Against Cancellation: an Indian Railways ticket that lets you board and share a side berth with one other passenger until a full berth frees up.
Race condition
A bug where the result depends on who gets there first. Example: two people both "win" the last seat because the system checked for both before it saved either.
RAG
Retrieval-augmented generation: first finding relevant passages in your own documents, then having an AI answer only from them.
Range query
A lookup for every entry between two values, like all orders from 1 to 7 March, rather than one exact key. Sorted storage makes it fast because matching entries sit next to each other.
Rank (in LoRA)
The width of the small add-on grids that LoRA trains, such as 8 or 64. A higher rank can learn more complicated changes but has more numbers to train and store.
Ranking
Putting search results in order, most useful first, using clues like how well a page matches your words, how trusted it is and how fresh it is.
Ranking score
A number the app works out for each item, guessing how likely you are to care about it. Items are then shown highest score first.
Rate limiting
Capping how many requests a client may make in a period, rejecting the excess.
RBAC
Role-based access control: permissions are given to roles, such as 'manager' or 'support agent', and each person gets the permissions of their roles.
Re-embedding
Working out every stored embedding again, usually after switching embedding models, because embeddings from different models can't be compared.
ReAct
An agent loop of reasoning about what to do, acting with a tool, observing the result and repeating until done. Short for 'reason and act'.
Read amplification
How many places a database has to look, or how many disk reads it makes, to answer one lookup. Checking five files to find one key means a read amplification of five.
Read receipt
A small message sent back when you open a chat, telling the sender you've seen their message (the blue ticks).
Read repair
Fixing an out-of-date copy during a normal read: when the copies asked disagree, the store returns the newest value and quietly updates the stale copy.
Read replica
A copy of a database that follows the primary's changes and serves reads, taking load off the primary.
Read your own writes
A promise that after you change something, you'll see your own change, even if others see it a little later.
Read-only
Data that people look at but nobody changes. It is the easiest kind to make fast, because copies of it can be kept in many places without ever going out of date.
Read-to-write ratio
How many times data is read for each time it's written. A news feed might be read 100 times per post; a GPS tracker writes far more than it reads.
Readiness check
A health check that asks whether a server can take real requests right now. If it fails, traffic is sent elsewhere, but the server isn't restarted, for example while it is still starting up.
Reading order
The sequence in which the blocks on a page should be read, such as down the left column before the right. Getting it wrong mixes up sentences and table rows.
Real-time factor
How long a system takes to process audio compared with the audio's length. A factor of 0.01 means one hour of audio is transcribed in about 36 seconds.
Real-time inference
Running a model at the moment a request arrives, so the answer uses the very latest information. It must finish within the request's time limit.
Real-user monitoring
Measuring speed and errors on users' own phones and browsers, so the team sees what people really experience, including slow networks and app crashes.
Reason code
A short code saved with an automatic decision saying why it was made, such as 'new device' or 'amount far above usual', so the decision can be explained later to a customer or an auditor.
Reasoning model
A model trained to work through a problem in hidden steps before it answers. Much better at multi-step problems, but slower and dearer per answer.
ReBAC
Relationship-based access control: deciding who may open something from relationships such as 'shared with', 'member of' or 'owner of', the way file sharing works in office apps.
Rebalancing
Reassigning which reader in a consumer group handles which partitions, whenever a reader joins, leaves or crashes. Reading pauses briefly while it happens.
Recall
Of all the items that should have been picked out, the share the system actually caught. In search it is often measured within the top few results, called recall at k.
Reciprocal rank fusion
A way to merge ranked lists from different searches using only each item's position, not its score. An item that ranks well in several lists rises to the top.
Recommendation
A suggestion an app picks for you, like a product, film or song, based on what you and other people have done before.
Reconciliation
Comparing two sets of records (yours and the bank's, say) to find and fix anything that doesn't match, like money that left one account but never arrived.
Recurrent network (RNN)
An older kind of language model that reads text one word after another, passing a running summary along. Early words fade by the end of a long text, and the work can't be split across many chips.
Red-teaming
Deliberately attacking your own system, the way a real attacker would, to find weak spots before someone else does.
Redaction
Hiding or replacing personal details, such as names, phone numbers or account numbers, before text is saved or shared.
Redirect
A response that says "what you want is at this other address", and sends the browser there automatically.
Redundancy
Keeping spare copies of important parts (servers, databases, even data centres), so one can fail without the app going down.
Reflection (agents)
An agent pausing to check whether a step or its final answer actually achieved the goal, and noting what went wrong so its next attempt is different.
Refresh token
A longer-lived key the app keeps after login and trades in for a fresh short-lived pass when the old one expires. The trade-in step checks the user is still allowed, so access can be cut off.
Region
A cloud provider's group of data centres in one city or area, such as Mumbai or Hyderabad. Regions are hundreds of kilometres apart, so one local disaster can't hit two of them.
Region pinning
Tying a piece of data, and every computation on it, to an allowed region such as India, and having the system enforce that rule on every path, including fallbacks and backups.
Regression test
A test that checks something which once worked, or once broke, still behaves correctly after a change.
Regular expression (regex)
A short pattern that describes a shape of text, such as "ten digits starting with 6 to 9", so a program can find every piece of text with that shape.
Reinforcement learning
Training by trial and reward: the model tries something, gets a score, and is nudged towards whatever scored well.
Rejection sampling
Letting a model try a problem many times, keeping only the attempts a checker confirms are right, and training on those. It is a common way to improve models at maths and code.
Release bundle
Everything a model needs to behave as tested, released and rolled back together: the model version, its threshold, its feature definitions and, for AI features, its prompt version.
Relevance ranking
Putting search results in order, best first, by scoring each one on how well it matches your words and on signals like sales, ratings and delivery speed.
Remote config
Settings that an app reads from a central service while it runs, so they can be changed for some or all users in seconds without shipping a new version of the app.
Repetition penalty
A sampling setting that lowers the odds of tokens the model has already used, the more often they've appeared, so long answers don't get stuck repeating the same words.
Replica (of a model)
One complete, working copy of a model on a group of GPUs that can answer requests by itself. A fleet runs several replicas behind a router and adds or removes them as traffic changes.
Replication factor
How many copies of each piece of data are kept, often 3.
Replication lag
The delay between a change on the primary database and that change appearing on a replica.
Request
A message asking a server to do something, like "send me this page" or "save this order".
Request coalescing
When many identical requests arrive together, sending only one of them to the source and giving all of them its answer.
Request fingerprint
A short code worked out from the important parts of a request, such as the amount and the order number. Comparing fingerprints quickly shows whether two requests ask for exactly the same thing.
Requests per second (rps)
How many requests arrive at, or are handled by, a system every second.
Reranker
A small AI model that reads a question alongside each result on a shortlist and reorders them, best match first. It is too slow to read everything, so it only checks the top few dozen.
Reservation chart
The final list of who has which berth on a train, prepared a few hours before departure. After it's made, waitlist positions stop moving.
Reserved capacity
Servers or GPUs booked for months or years at a lower hourly rate. You pay whether they are busy or idle, so it suits work that keeps them busy all day.
Reset token
A long random code inside a password reset link. It proves you opened the email, works only once, and expires after a short time.
Resolution rate
The share of support chats where the customer's problem was actually solved, usually counted as no repeat contact about the same issue within a day or two.
Response
The server's answer to a request: the data, a confirmation, or an error.
Response buffering
When a server, proxy or CDN collects a whole reply (or a large block of it) before passing it on. Efficient for ordinary web pages, but it turns a streamed AI answer into one lump at the end.
REST
A common style of API where every thing has an address (like /orders/812) and HTTP methods are the verbs.
Resumable upload
Sending a big file in small pieces, so if the connection drops, only the missing pieces are sent again instead of starting over.
Retention period
How long an event log keeps events before deleting the oldest ones, for example seven days. Readers must catch up within it or miss events.
Retraining
Training a model again on newer examples so it keeps up with a changing world, then checking it beats the current model before it replaces it.
Retry
Trying again after something failed or took too long. Safe retries make sure the action still happens only once.
Retry budget
A cap on how many retries a system may send, often a small share of normal traffic such as 10%. It stops retries from piling extra load onto a service that is already failing.
Retry storm
A flood of retries from many clients at once, often right after a brief failure, that knocks a recovering service over again. Random jitter and a cap on retries prevent it.
Revenue management
How airlines and hotels decide how many seats or rooms to sell at each price, and change those numbers as bookings come in, to earn the most from each flight or night.
Reversal
Putting money back after a payment fails partway: the debit is undone and the amount returns to your account.
Reverse geocoding
Turning a GPS location, a pair of numbers, into a place name people understand, such as "Calangute, Goa".
Review queue
A list of items waiting for a person to check, such as unsure extractions or flagged posts, usually sorted by priority.
Reviewer agreement
How often two trained people, looking at the same case on their own, reach the same decision. Low agreement means the case or the rule is genuinely unclear.
Reward
The number a learning system tries to increase after each choice, such as 1 if the person watched for 10 minutes and 0 if not. What you choose as the reward decides what the system learns to do.
Reward hacking
When an AI system finds ways to score highly with its scorer without actually getting better, such as writing longer, gushing text because the scorer happens to like it.
Reward model
A separate model trained to score how good an answer is, standing in for human judgement during training.
RFID
Radio-frequency identification: a small tag with no battery that answers a nearby reader's radio signal with its ID number. FASTag stickers and shop security tags work this way.
Risk assessment
A written check, done before launch and after big changes, of who a system could harm, how likely and how badly, and what is being done to reduce it.
Risk score
A number that says how unusual a sign-in looks compared with your normal pattern. The higher it is, the more checks the service asks for.
RLHF
Reinforcement learning from human feedback: people compare pairs of answers, a reward model learns their preferences, and the main model is trained to score well with that reward model.
RLVR
Reinforcement learning with verifiable rewards: training on problems whose answers a program can check, such as maths with known answers or code that must pass tests. It is the main way reasoning models are trained.
Road graph
A map stored the way a computer can search it: junctions as points, joined by road segments that know their length, direction and allowed turns. Route planners search it for the quickest chain of segments.
Road segment
A short stretch of road, often a few hundred metres between two junctions. Maps apps judge traffic one segment at a time.
robots.txt
A small text file a website puts at its root to tell crawlers which parts they may visit and how often. Polite crawlers read it first and obey it.
Rollback
Switching back to the previous working version when a release causes problems.
Rolling deploy
Updating servers a few at a time, each joining traffic only after passing its health check, so the app never goes down.
Rolling summary
A short summary of a conversation's older messages, updated as the chat grows, that is sent to the model in place of those messages.
Rolling window
A limit counted over the most recent stretch of time, such as the last 60 seconds or the last 24 hours, instead of resetting for everyone at a fixed clock time.
Round robin
Sharing work by taking turns: the first request goes to server 1, the next to server 2, and so on, then back to the start.
Round trip
A message going to another computer and the answer coming back. Its time is the wait you actually feel.
Router
A step that sorts each incoming request and sends it down the path best suited to it, such as a particular prompt, tool, agent or model.
Routing engine
Software that works out the real road path and travel time between two points, including one-way streets, dividers and U-turns.
Row-level security
A database feature that adds a rule to every query, such as 'only rows for this company', so even code that forgets to filter can't read other customers' data.
RPO (recovery point objective)
How much recent data a business agrees it could lose in a disaster, measured in time. An RPO of 5 seconds means losing up to the last 5 seconds of changes is acceptable.
RTO (recovery time objective)
How long a service is allowed to be down after a disaster before it must be serving users again, such as 15 minutes.
Rubric
The list of areas an interviewer scores you on, each with a scale and notes on what a strong answer looks like.
Rules engine
Software that applies a list of written business rules, such as 'hold large payments from new devices', alongside or on top of a model's score.
Runbook
A short written guide for a specific alert or problem: what it means, what to check first, and the safe steps to fix or contain it.
Saga
A long action split into small steps, each saved on its own, where every step has a matching undo. If a later step fails, the earlier ones are undone in reverse order.
Salt
A random extra bit mixed into each password before it's scrambled, different for every person, so two people with the same password end up with different fingerprints.
Sample size
How many users or orders an experiment needs in each group before a difference between the groups can be trusted rather than put down to luck.
Sampling
Showing or handling only some of a very large number of things, like a few comments a second out of thousands, when all of them would be too many.
Sampling (AI)
Picking the next token from the model's list of probabilities, sometimes choosing a less likely one. It is why the same question can get different answers.
Sanchar Saathi
A Government of India portal where you can see the mobile connections registered in your name, report ones you didn't take, and block a lost or stolen phone.
Sandbox
A sealed-off space where untrusted code can run without reaching real data, other programs or the open internet.
Scalar quantisation
Storing each number in a vector with fewer bits, such as one byte instead of four, so the vectors take a quarter of the memory with only a small loss in accuracy.
Scaling law
A rule found by experiment for how a model's quality improves as its size, its training text and the compute spent on it grow. Labs use it to plan how big a model to train for a given budget.
Scatter-gather
Answering a question by sending it to every shard or server at once, then merging their answers. It works, but it's as slow as the slowest one.
Schema
A description of the shape data must have: which fields exist, what type each one is and which are required. Programs use it to check data before trusting it.
Schema migration
A planned change to the shape of a database, such as adding a column, renaming one or adding an index, applied to the live data.
Schema registry
A shared service that stores the agreed format of each kind of event and rejects changes that would break the readers already relying on it.
Schema validation
Checking data in code against its schema, field by field, and rejecting it if anything is missing, the wrong type or outside the allowed values.
Scope (permission)
One named thing an app asks to be allowed to do with your account, such as see your email address or save files to your Drive. The consent screen lists them.
Scoped credential
A key or token limited to certain actions and data, and often to a short time, so if it leaks or is misused the damage stays small.
Search log
A record of what people searched for and what they clicked or bought next. Shops use it to learn common misspellings and which results people really wanted.
Search query
The words you type into a search box. The system cleans them up and works out what you mean before looking anything up.
Seasonality
Regular, expected swings in behaviour tied to the calendar, such as salary day, weekends, IPL evenings or Diwali shopping. They look like drift but come back every time.
Seat map
The picture of a hall or coach showing which seats are free, held or booked, updated as other people choose.
Second factor
A second proof that it's you, on top of your password: usually a code or prompt on your phone. Also called two-step verification or 2FA.
Secrets vault
A locked store for passwords and keys that hands them out only to approved programs, often for a short time, and records every use. With AI agents, it lets tools use a key without the model ever seeing it.
Security code
A long number (WhatsApp shows 60 digits) worked out from the keys in a chat. If you and your friend see the same code, nobody has swapped in a fake key.
Seed (random seed)
A starting number for a random process. Using the same seed repeats the same 'dice rolls', so some AI services accept one to make answers more repeatable, though not guaranteed identical.
Selection bias
When the data you collect comes from an unusual slice of people or cases, such as only very unhappy users bothering to rate, so it gives a distorted picture of everyone.
Self-consistency
Asking the same question several times with some randomness and taking the most common answer.
Self-correction
An AI agent reading the error or result from a failed step and changing its plan, for example fixing a wrong input or trying a different tool, instead of stopping or repeating itself.
Self-hosting
Running a model on your own or rented servers instead of paying a provider per call. You gain control but take on GPUs, scaling and updates.
Semantic cache
A cache that reuses an earlier AI answer when a new question means the same thing, skipping the expensive model.
Semantic Kernel
Microsoft's open-source AI framework for C#, Python and Java, which turns a company's existing functions into plugins a model can call. Microsoft now points new work to its successor, the Microsoft Agent Framework.
Semantic memory
General facts an assistant has learned about a person or the world, such as 'is vegetarian', kept apart from when and where they were learned.
Sender check
Checking that an email really came from the domain it claims, by looking up the domain's published list of allowed servers and verifying its digital seal. Also called email authentication.
Sender reputation
A running record of how a sender's past emails were received: how many were opened, ignored or reported as spam. Good senders get through easily; new or badly behaved ones are treated with suspicion.
Sequence number
A number that goes up by one with every update, so an app can put updates in order and ignore old ones that arrive late.
Server
A computer (or program) that waits for requests and answers them.
Server-sent events (SSE)
A way for a server to keep sending small updates over one open web request, used to stream AI replies to the screen as they are written.
Session
The server-side record that you are logged in, created when you sign in and checked on every request.
Session token
A long random key your app gets after you log in and sends with every request, so you stay logged in without a new code each time.
Settlement
The actual movement of money between banks after payments or refunds are approved. It usually happens in scheduled batches, once or a few times a day, not one payment at a time.
Shadow AI
AI tools that staff use for work without the company approving or knowing about them, such as pasting customer data into a free public chatbot.
Shadow traffic
Sending copies of real requests to a new version in the background and comparing its answers, without users ever seeing them.
Shard
One piece of a database that has been split up. Each shard is a separate database holding part of the data, like users A to H.
Shard key
The piece of data (often the user ID) used to decide which shard a record lives on.
Sharded data parallelism
Data parallelism where the weights, gradients and optimiser state are split across the GPUs instead of copied onto each one, and pieces are fetched over the network just when needed. ZeRO and FSDP are well-known versions.
Sharding
Splitting one large dataset across several databases, each holding a part.
Showback
Showing each team or feature what its usage costs, without billing it for that cost. It builds awareness before budgets or internal charges are set.
Shrinkage
The share of a paid shift that isn't spent on the actual work: breaks, training, meetings, quality reviews and leave. Staffing plans add it back on top of the raw hours.
Side effect
A change an action makes in the outside world, such as sending an email or charging a card, which can't be undone just by running the code again.
Signed link
A link to a private file that carries a tamper-proof stamp from the app and stops working after a set time, often a few minutes. Anyone who finds it later gets nothing.
SIM swap
A fraud where someone tricks or bribes a mobile operator into moving your number to a new SIM card they hold, so your OTPs and calls go to them.
Similarity threshold
The cut-off for how alike two embeddings must be before they count as a match. Set it too loose and different questions match; too strict and almost nothing does.
Simulated user
An AI model that plays a customer or user during testing, following a script of goals and quirks, so a chat agent can be tried on hundreds of conversations before real people use it.
Single point of failure
One component whose failure takes the whole system down.
Single sign-on
Signing in once with one account, such as Google, and then being let into many apps without typing a password for each.
Sitemap
A file a website publishes listing its pages, often with when each last changed, so crawlers can find and refresh them without guessing.
Skill-based routing
Sending each item in a queue only to people who have the skills it needs, such as a language, a policy area or a training level, instead of to whoever is free.
SKU
Stock keeping unit: one exact version of a product that a shop counts on its own, such as the blue 128 GB model of a phone.
SLA (service level agreement)
A promise written into a contract with customers, with a penalty such as refunds or credits if it is missed. It is usually looser than the team's own internal target.
SLI (service level indicator)
The measured number behind a reliability goal, such as the share of payments that succeed or the share of pages that load within 2 seconds.
Sliding window
A way of counting over the most recent stretch of time, such as the last 60 seconds, measured back from now, so the window moves forward as time passes.
Sliding-window attention
A cheaper form of attention where each token looks only at a fixed number of recent tokens, say the last 4,000, instead of the whole text. Models mix it with a few full-attention layers to keep long-range recall.
SLO
Service level objective: a target for something users feel, like "99.9% of payments succeed each month".
Sloppy quorum
Counting a stand-in machine towards the quorum when a usual copy is down, so writes keep working. It keeps the app up but loses the guarantee that reads always see the latest write.
Slow hash
A way of scrambling passwords that is deliberately slow. You never notice the delay when you sign in, but someone trying millions of guesses has to wait for every one.
Small language model
A language model small enough to run cheaply or on a phone or laptop, usually trained or distilled for narrower jobs.
SMS gateway
A service that sends text messages for an app by handing them to mobile networks like Jio, Airtel or Vi.
SMS pumping
A fraud where bots trigger huge numbers of OTP texts to numbers that pay the fraudster a share of each SMS, leaving the app with the bill.
Snapshot
A full copy of something saved at one moment, like a photo of a document at 11:50. Version history is a list of snapshots you can go back to.
Social engineering
Tricking a person, not a computer: using fear, urgency or a fake identity on a call or message so the victim hands over money or details themselves.
Social graph
The stored record of who follows or is friends with whom. Apps look it up to know whose feeds a new post should reach.
Soft delete
Hiding something straight away so nobody can see it, and actually removing it later. It makes deleting quick and lets mistakes be undone.
Soft labels
A teacher model's full set of probabilities for every possible answer, not just its final pick. They show a student model which wrong answers were nearly right, which makes for a richer lesson.
Sorted file (SSTable)
A file an LSM-tree database writes once, holding keys in sorted order, and never changes afterwards. Short for sorted string table. Newer files hold newer versions of a key.
Source of truth
The one place whose data counts as correct. Other copies may be a little behind, but when they disagree, this one wins.
Source reputation
A score for how far a website or account can be trusted, built from signals such as its age, its history and how often it has spread spam or false claims.
Space amplification
How much more disk a database uses than the live data itself needs, because of half-empty pages, old versions or deleted data that hasn't been cleaned up yet.
Spam filter
Software on the mail server that checks every incoming email and decides whether it goes to your inbox, your Spam folder, or nowhere at all.
Spam score
A single number that combines everything the filter found out about an email. Below a threshold it goes to the inbox; above it, to Spam or straight back to the sender.
spam trap
An email address that no real person uses, kept only to catch spam. Nobody ever signed it up for anything, so any bulk mail that reaches it gives the sender away.
Span
One timed step inside a trace, such as a search or a model call, recorded with its start, end and details.
Speaker diarisation
Working out who spoke when in a recording, so each line of a transcript is labelled with the right speaker, such as doctor or patient.
Speculative decoding
A small, fast model guesses the next few tokens and the big model checks them all in one step, keeping the ones it agrees with. The answer is the same, only quicker.
Speech to text
Software that listens to spoken words and writes them down as text. Also called speech recognition or ASR.
Speech-to-speech model
An AI model that hears spoken audio and answers directly in spoken audio, without writing the words down as text in between.
Split brain
When two machines both believe they are the leader and both accept changes, so the data stops agreeing.
Spot instance
Spare cloud capacity sold cheaply that the provider can take back at short notice. It suits work that can pause and restart, such as training.
Spotlighting
Clearly marking untrusted text in a prompt, for example by wrapping or encoding it, so the model treats it as data rather than as instructions.
SQL database
A relational database: data in tables with fixed columns, queried with SQL, with strong guarantees about consistency.
SSD
A solid-state drive: storage made of memory chips with no moving parts. It keeps data when the power is off and reads much faster than an old spinning hard disk, though still far slower than RAM.
Stale data
A copy that is out of date because the original has changed since it was made.
STAR method
A simple shape for telling a work story in an interview: the situation, your task, the actions you took and the result. Many people add a fifth part, what you learned.
State machine
A fixed list of stages something moves through, in order. An order goes placed → accepted → cooking → picked up → delivered, never backwards.
Stateless server
A server that keeps no user-specific memory between requests, so any copy of it can answer any request.
Static batching
Running a fixed group of requests together from start to finish. The group ends only when its longest answer ends, so places freed by short answers sit empty.
Static page
A web page that is exactly the same for every visitor, so it can be made once and handed out from a CDN instead of being built afresh for each person.
Statistical significance
A check that a difference between two groups is big enough, for the amount of data, that it is unlikely to be plain luck. The usual bar allows about a 1 in 20 chance of a false alarm.
Status code
A three-digit number in every response that says how it went: 200 means OK, 404 means "no such page", 500 means the server had a problem.
Status page
A public web page where a company says which of its services are working, which are having problems, and what it is doing about them.
Stemming
Cutting words back to a shared root before searching, so "shoes", "shoe" and "shoeing" are all treated as "shoe" and find the same products.
Step budget
A limit on how many steps or tool calls, or how much money, an agent may use on one task before it must stop.
Step-up check
An extra check asked for only when something looks unusual, like an OTP or a 'Is it you?' prompt on a phone you already use.
Sticky assignment
Putting each user in the same experiment group every time, on every device, usually by working out the group from their user ID instead of picking at random on each visit.
Sticky session
A load balancer setting that keeps sending the same user to the same server, usually by a cookie. It lets that server keep things in memory, but unbalances load and loses them if the server dies.
Stock bucket
A slice of one product's stock kept in its own counter, so a rush of buyers is spread over several counters instead of all hitting one.
Stop condition
A rule, written in the code around an AI agent, that ends a run: the task is verified as done, it has clearly failed, or it has hit a limit on steps, time or money.
Stop reason
The label an AI service attaches to each reply saying why the model stopped writing: it finished, it hit the length limit, it reached a stop sequence, or it refused. Apps check it before trusting the reply.
Stop sequence
A piece of text, such as '###', that tells the model to stop writing as soon as it produces it. Apps use it to end answers cleanly at a known point.
Stop-sell
An instruction from a hotel telling every site to stop offering a room type for certain nights, even if their copies still show rooms left. Hotels use it when they're nearly full.
Storage engine
The part of a database that decides how data is laid out on disk and in memory, and how it is written and found again. Two databases with the same query language can behave very differently because of it.
Store and forward
Holding a message safely until the receiver is reachable, then delivering it. How chat apps deliver messages sent while your phone was off.
Straggler
One machine or GPU running slower than the rest. When every step waits for all GPUs to finish, a single straggler slows the whole cluster to its pace.
Straight-through processing
Handling a document or transaction from start to finish with no person touching it. The share handled this way is a key measure of how well an automated system works.
Stratified sampling
Picking examples group by group, such as by task type or language, so rare groups get enough examples instead of being drowned out by common ones.
Stream delay
How far a live stream on your screen runs behind the real event. On a normal live stream it is often somewhere between 10 and 30 seconds.
Stream processing
Computing results continuously as events arrive (counts per minute, fraud scores) instead of in nightly batches.
Streaming
Sending something in small pieces as soon as each piece is ready, instead of waiting until the whole thing is finished.
Strong consistency
A promise that every read sees the latest saved change, as if there were only one copy of the data, even though there are several.
Structured log
A log line written as named fields (time, request ID, user, error) rather than free text, so it can be searched and filtered precisely.
Structured output
Asking an AI model to reply in a fixed shape, with named fields, instead of free text, so a program can read the answer directly.
Student model
The smaller, cheaper model trained to copy a teacher model's answers for one particular job.
Sub-agent
A helper agent started by another agent for one sub-task. It works in its own fresh context and hands back only a short result.
Subscriber database
The phone network's list of every SIM allowed to connect, with its number and what it may do (calls, data, roaming). A SIM that isn't on it can't join the network.
Sufficiency check
A quick check, usually a small AI model call, that asks whether the passages found so far are relevant and enough to answer the question. It decides whether the assistant answers now or searches again.
Supply-chain attack
An attack through something you depend on, such as a downloaded model, software package or plugin, rather than through your own code.
Surge pricing
Prices that rise in an area when many people want rides and few drivers are free, so more drivers head there and demand cools down.
Sync
Bringing two copies of the same thing up to date with each other, like the cart saved on your phone and the one on the shop's servers.
Sync conflict
When two copies of the same data were changed differently while apart, so the system has to decide how to combine them.
Synchronous replication
Copying each change to a replica and waiting for it to confirm before telling the app the save worked. Nothing confirmed is lost if the primary dies, but every save takes a little longer.
Syntax tree
A structured map of a code file that a program builds by parsing it, showing its functions, classes and how the pieces nest, so tools can find and link code exactly.
Synthetic data
Training or test examples written by an AI model rather than collected from real people, then filtered for quality and variety.
Synthetic monitoring
Robot checks that act like a user from outside the system, such as loading the home page or making a test payment every minute, to catch problems even when real traffic is quiet.
System design interview
A job interview round, usually 45 to 60 minutes, where you design a large app out loud on a whiteboard. It tests how you reason, not whether you remember one right answer.
System prompt
The fixed opening instructions an app gives an AI model on every call: its role, rules and tone.
System prompt leakage
A model revealing its hidden instructions, and any secrets wrongly placed in them, when users ask cleverly.
Table scan
Reading every row of a table to find the few that match. Fine for small tables, painfully slow for big ones.
Tamper check
Looking for signs that an ID document photo was edited or faked, such as a swapped photo, changed digits, mismatched fonts or a screenshot instead of a real card.
Target leakage
When a training feature quietly contains information about the answer the model is meant to predict, so it scores brilliantly in testing and poorly in real use.
Task decomposition
Breaking a big goal into smaller sub-tasks, and those into smaller ones if needed, until each piece is simple enough to do in one step or a short loop.
TCP
A careful way to send data: it numbers every packet, checks they all arrived in order, and re-sends any that went missing.
Teacher model
The large, capable model whose answers are used to train a smaller model in distillation.
Telephony
The phone network side of a call: carrying a phone call between a caller's phone and a computer system, with its own delays and audio quality limits.
Temperature
A setting that controls how adventurous the model's word choices are. Low gives predictable, focused answers; high gives more varied ones that are also more likely to wander.
Template variable
A named blank in a prompt template, such as the customer's name or order number, that the app fills in for each request before the prompt is sent to the model.
Tenant
One customer company using a shared software product. Its data lives in the same system as other customers' but must stay strictly its own.
Tensor parallelism
Splitting each layer of a model across several GPUs that work on every token together. It lets a model too big for one GPU run fast, but needs very fast links between the GPUs.
Terminal
A text window where developers type commands to run programs, tests and tools, instead of clicking buttons.
Test-time compute
Spending more calculation while answering, such as longer thinking or several attempts, to get better results without a bigger model.
Text layer
The real, selectable text stored inside a PDF made by software. When it is present and sensible, an app can read the page exactly without any OCR.
Text normalisation
Cleaning text into a standard form before checking it, such as turning 'h@te' into 'hate' and removing invisible characters, so simple disguises don't slip past filters.
Text to speech
Software that reads text aloud in a natural-sounding voice. Also called speech synthesis or TTS.
Thermal throttling
A chip slowing itself down when it gets too hot, to protect itself and the device. Long AI jobs on a phone can trigger it, so answers get slower.
Thinking budget
A cap on how much a reasoning model may think for one request, trading answer quality against time and cost. Some providers call it reasoning effort.
Thinking tokens
The working-out a reasoning model writes before its final answer. They are usually charged for even when you can't see them.
Three-way match
An accounts check that an invoice agrees with the purchase order and the delivery note on items, quantities and prices before the bill is paid.
Throttling
Deliberately sending or handling something less often than it happens, like sharing a moving cursor's position a few times a second instead of on every tiny movement.
Throughput
How much work a system finishes per unit of time, such as requests per second.
Thumbnail
A small, low-detail version of a photo or video, used in grids and previews because it loads almost instantly.
Time decay
Counting older events for less than recent ones, so a search from today matters more than one from last month and yesterday's craze fades on its own.
Time to first token
How long an AI model takes to send back the first piece of its answer. It decides how quickly a reply seems to start, even when the full answer takes longer.
Time to ready
How long a new server takes from being requested to serving its first real answer: finding a machine, loading the model and warming up. For big AI models it is often several minutes.
Time window
A fixed stretch of time, such as 7:20 to 7:25, whose events are added up together before the next stretch begins. Counting in windows turns a flood of events into a few totals.
Time-based split
Splitting past data by date: train the model on older data and test it on a later period, the way it will really be used. It stops the model from learning from the future.
Time-series data
Measurements recorded again and again over time, like a phone's battery level every minute or a shop's sales every hour. Mostly new rows are added; old ones are rarely changed.
Timeout
A limit on how long to wait for something before giving up and reporting an error.
Timestamp
The date and time attached to a piece of data, saying when it was made. Comparing timestamps tells a system which of two updates is newer.
Timing side channel
A leak where an attacker learns something secret from how long an operation takes rather than by reading the data, such as noticing that a cached prompt gets a faster reply.
TLS
The encryption that secures web and app traffic, shown by the padlock in a browser. The two sides agree a secret key with a short handshake before any data is sent.
TLS termination
Decrypting a secure (HTTPS) connection at the front door, such as a gateway or load balancer, so the servers behind it receive plain requests and don't each have to do the unlocking.
Token (AI)
A small piece of text, roughly a word or part of one, that an AI model reads and writes one at a time.
Token bucket
A way to rate limit: each user has a bucket of tokens that refills steadily; every request uses one, and an empty bucket means "wait".
Token vault
A locked, tightly controlled store that maps placeholders back to the real personal values they stand for, with every lookup recorded and entries deleted on a schedule.
Tokenisation (of personal data)
Swapping a personal value for a placeholder such as [PHONE_1] and keeping the real value in a locked store, so it can be put back later only where it is needed.
Tokeniser
The part of an AI system that cuts text into tokens using a fixed vocabulary. Common English words become one token, while numbers, rare words and many Indian-language words split into several.
Tokens per minute
A limit a model provider sets on how many tokens a whole account may send and receive each minute, counted across all of its users together.
Tokens per second
How fast a model writes once it has started answering. Together with time to first token, it describes how quick a reply feels.
Tombstone
A marker saying "this was deleted at this time", stored instead of simply erasing the data, so every copy learns about the delete and the old value isn't brought back by repairs.
Tool calling
A way for an AI model to ask the app to run a function it has been told about, by replying with the function's name and its inputs in a fixed format.
Tool poisoning
Hiding instructions for the AI model inside a tool's name or description, so that simply installing a tool can steer the model into leaking data or misusing other tools.
Top-k
A sampling setting that keeps only the k most likely next tokens, say 40, and picks among those, so very unlikely tokens are never chosen.
Top-p (nucleus sampling)
A setting that keeps only the most likely next tokens whose probabilities add up to a chosen share, such as 90%, and picks among those.
Topic
A named stream of events inside an event log, such as "orders" or "payments". Producers write to it and readers subscribe to it.
Trace ID
A unique ID attached to a request when it enters the system and passed along to every service it touches, so all its logs and timings can be found together.
Trade-off
A choice where getting more of one thing means getting less of another, such as speed against freshness or cost against reliability.
Training data
The examples a model learns from. For fine-tuning, these are pairs of inputs and the exact answers you want, checked by people.
Training-serving skew
When a model's inputs are worked out one way while it learns and a slightly different way when it answers live, so it does worse in real use than in testing.
Trajectory
The full sequence of steps an agent took on a task: its reasoning, tool calls and results. Checking it shows whether the path was sensible, not just whether the outcome was right.
Transaction (in a database)
A group of changes the database treats as one: either all of them are saved or none are. Saving an order and its items together is a transaction, so you never end up with half an order.
Transaction ID
A unique reference number for one payment. Every system involved uses it, so the payment can be tracked, checked and never done twice.
Transcoding
Converting a video into other sizes and qualities (say 1080p, 720p and 360p) so every phone and connection can play a version that suits it.
Transformer
The neural-network design behind almost every modern language model. Its key ingredient is attention, which lets each token draw on all the others at once.
Transient error
A failure that is likely to go away by itself, such as a timeout or a busy server. Worth a short wait and a retry, unlike errors such as 'not found' that will repeat every time.
Transliteration
Writing a word from one language in the letters of another, like typing the Hindi word "chappal" in English letters.
Tree of thought
A way of solving hard problems where the model proposes several possible next steps, each is scored, weak ones are dropped and the best ones are explored further, instead of following a single line of reasoning.
Trie
A tree of letters where each step down adds one character, so every word is a path from the top. Autocomplete stores the most popular endings at each point to answer a prefix instantly.
TTL
Time to live: how long a cached copy is trusted before it expires and is fetched again.
Tumbling window
A way of counting in fixed boxes of time that sit end to end without overlapping, such as 10:00 to 10:01, then 10:01 to 10:02. Each event lands in exactly one box.
Two-phase commit (2PC)
A way to make changes in several databases all or nothing. A coordinator first asks everyone to get ready and lock what they need, then tells them all to commit, or all to cancel.
Two-tower model
A recommendation model with two halves: one turns a person's history into a list of numbers, the other does the same for each item, so items whose numbers sit close to the person's can be found quickly.
Type check
An automatic check that every piece of code is used with the right kind of data, such as a number where a number is expected. It catches many mistakes before the code runs.
Type safety
Declaring the exact shape of data (this field is a number, that one a date) so the code checks it automatically and mistakes are caught early instead of surfacing later as odd bugs.
UDP
A fast way to send data with no checking: packets just go, and a lost one is simply skipped. Used for live calls and games.
UIDAI
The Unique Identification Authority of India, the government body that runs Aadhaar and answers requests to confirm that a person is who they say they are.
Unit economics
The cost and the revenue for one unit of a business, such as one conversation or one paying user, which shows whether growth makes or loses money.
Unsafe output handling
Passing a model's output straight into another system, such as a database query, web page or command line, without checking it, so injected text can become an attack.
UPI ID
An address like name@bank that points to a bank account, so people can pay you without knowing your account number. Also called a VPA (virtual payment address).
Upscaling
Turning a small or low-resolution image or video into a larger, sharper one, often with an AI model filling in fine detail.
URL
The full web address you see in the browser's address bar. It names the website and the exact page or file you want, for example https://shop.example/diwali.
URL frontier
A crawler's to-do list of web addresses to visit, split by website so each site is visited at a polite pace.
URL normalisation
Cleaning up a web address into one standard spelling, for example lowercasing the site name and removing tracking tags, so the same page isn't treated as many different pages.
Usage limit
How much of a service one person or plan may use in a set time, such as 40 messages or a number of AI tokens a day. It resets at a known time.
Usual (historical) speed
How fast traffic normally moves on a road at a given hour and day of the week, worked out from weeks of past data.
Utilization
The share of time a component is busy. Near 100%, queues grow and latency climbs fast.
UUID
A long random-looking ID, such as 3f9a1c2e-…, that any machine can create on its own without clashing with others. Fully random ones scatter new rows all over an index.
Validation and repair loop
Checking a model's output against rules or a schema and, if it fails, sending the errors back to the model to fix, a limited number of times.
Validation set
A slice of examples kept apart from the practice data and used to choose between candidate prompts or models. It is separate from the final test set, which is used only once to report the real score.
Vector clock
A small list of counters, one per machine or writer, attached to each version of data. Comparing two lists shows whether one version came after the other or whether they were changed separately and clash.
Vector database
A database built to find the stored embeddings closest to a given one, which means searching by meaning.
Vector extension
An add-on that lets an ordinary database, such as PostgreSQL with pgvector, store embeddings and run similarity search beside the rest of your data.
Velocity check
A check on how fast something is happening, such as five payments to new payees in ten minutes, which can be a sign of fraud.
Vendor lock-in
Being so tied to one supplier's product that switching to another would take a lot of time and money, which leaves you exposed to its price rises and outages.
Verifier
A program that checks whether an AI's answer is right, for example by comparing a maths result with the known answer or running code against its tests. Reasoning models are trained on its right-or-wrong marks.
Version control
A system that keeps every past version of a set of files, with who changed what and when, so a team can compare versions and go back.
Versioned key
A key that includes which batch or model version made the value, such as recs:b1021:user42, so new results never overwrite old ones in place and each version can be switched to or deleted as a whole.
Versioning (storage)
A storage setting that keeps older versions of a file for a while when it's replaced or deleted, so mistakes can be undone.
Vertical scaling
Handling more load by moving to a bigger machine (scaling up).
Video KYC
Checking a customer's identity on a live video call with a trained official, who sees their face and ID document and asks questions, instead of meeting in person.
Video segment
A small piece of a video, a few seconds long. Streams are sent as a series of segments so playback can start quickly and adjust as it goes.
Virtual key
A key an AI gateway gives each team or feature in place of the real provider key. The gateway tracks spending and limits per virtual key, and can switch one off without touching the real key.
Virtual node
One of many points a single machine is given on a hash ring. Owning lots of small slices instead of one big one spreads keys evenly and shares out the work when machines join or leave.
Virtual waiting room
An online queue. Instead of everyone hitting the booking system at once, people wait on a holding page and are let in a few at a time.
Vision transformer
An image model that cuts a photo into small square tiles and lets every tile compare itself with every other tile, the same idea that powers language models, applied to pictures.
Voice activity detection
Software that tells speech apart from silence and background noise, used to decide when someone has started or finished talking.
Voice cloning
Making a synthetic voice that sounds like a particular person from recordings of them. It should only be done with that person's consent.
VPN
Virtual private network: a service that sends your internet traffic through another computer, so websites see that computer's address and location instead of yours.
Waitlist
An ordered line of people waiting for something that's full, like train berths. When a place frees up, the person at the front gets it and everyone else moves up.
Wake word
The name a voice assistant listens for before it starts paying attention. It is spotted on the device itself, so nothing is sent until it is heard.
Warm pool
Spare servers kept running with the model already loaded, so extra traffic can be served without waiting minutes for new machines to start.
warming up
Starting a new email sender slowly, with a few hundred emails a day growing over weeks, so mail services can build up trust in it before it sends at full volume.
Watermark
A mark added to a picture, video or sound to show where it came from. AI tools often hide it in the pixels themselves, invisible to people but readable by software.
Watermark (in streams)
A stream job's running estimate that it has now seen nearly every event up to a certain time, which tells it when a time window can be treated as finished.
Web search API
A paid service that takes a search query and returns a list of matching web pages with titles, links and short snippets, so an app doesn't need its own copy of the web.
Webhook
A message one company's server sends to yours when something happens, like a bank telling you a payment succeeded.
WebSocket
A long-lived, two-way connection between client and server, used for live updates like chat or delivery tracking.
WebVTT
A simple text format for captions on the web. Each caption line comes with the time it should appear and the time it should disappear.
Word error rate
The share of words a speech-to-text system gets wrong, counting words it swaps, drops or adds. Lower is better, and it should be measured on audio like your real callers'.
Worker
A background process that takes jobs from a queue and does them.
Workflow (AI)
A fixed sequence of steps, some of them model calls, laid out in code ahead of time. More predictable and easier to debug than an agent choosing its own steps.
Workflow engine
Software that runs long multi-step processes reliably, saving progress after every step, retrying failures and resuming after crashes. Temporal is a well-known example.
Workflow versioning
Keeping track of which version of a long-running process's code each run started on, so runs already in progress can finish safely after the code changes.
Working memory
What an AI assistant has in front of it right now: the current conversation and task in its context window.
Write amplification
How many bytes a database really writes to disk for each byte of data you change. Rewriting an 8 KB page to change a 100-byte row, or rewriting files during compaction, both raise it.
Write-ahead log
A numbered, ever-growing list of every change a database makes, written down before the change itself. It lets the database recover after a crash, and replicas follow it to copy every change in order.
Write-back
Saving changes to the cache first and passing them to the database later in batches. Writes are very fast, but changes not yet passed on are lost if the cache fails.
Write-through
Saving every change to the cache and the database at the same time, so the cached copy is always up to date, at the cost of slightly slower writes.
Zero data retention
An arrangement where a model provider doesn't keep your prompts or answers once a request is finished.
Zero-shot prompting
Asking a model to do a task with instructions only and no worked examples.