AI glossary
AI glossary: the words of AI engineering in plain English
48 terms and 10 side-by-side comparisons, from tokens to MCP, each in one clear sentence and linked to the free lesson that teaches it hands-on.
48 terms10 comparisons
Terms A to Z
A
- Agent loop
- The agent loop is the cycle at the heart of an agent: send the conversation to the model, run any tool it asks for, add the result, and repeat.
- AI agent
- An AI agent is a program in which a language model chooses its own next step, calling tools in a loop until the task is done or it is stopped.
- Attention
- Attention is the step in a transformer where each token scores the tokens it may see, turns the scores into weights, and takes a weighted mix of their values.
C
- Caching
- Caching keeps the result of an earlier model call or computation so the same request can be answered again without paying for it twice.
- Canary rollout
- A canary rollout sends a small share of real traffic to a new model or prompt first, and widens it only if quality and errors stay within agreed limits.
- Chunking
- Chunking is cutting long documents into smaller passages before embedding them, so a search can return just the part that answers a question.
- Coding agent
- A coding agent is a model that reads a code repository, edits files and runs commands, such as the tests, in a loop until it has finished a task.
- Context engineering
- Context engineering is choosing what goes into a model's context on each call, instructions, history, documents and tool results, so it has what it needs and little else.
- Context window
- The context window is the most text, counted in tokens, that a model can take in one call: the prompt, the history and its own reply together.
E
F
- Few-shot prompting
- Few-shot prompting puts a handful of worked examples in the prompt so the model copies their pattern, format and judgement on a new input.
- Fine-tuning
- Fine-tuning is training an existing model further on your own examples so it learns a task, style or format it did not do well before.
G
- Golden set
- A golden set is a fixed collection of inputs with agreed good answers, used as the trusted reference an AI feature is scored against.
- Guardrails
- Guardrails are the checks around a model call that stop unsafe or unwanted input and output, enforced in code rather than only asked for in the prompt.
H
- Hallucination
- A hallucination is an answer from a language model that sounds confident and fluent but is not supported by its sources or is simply false.
- Hybrid search
- Hybrid search runs a keyword search and a vector search on the same query and merges the two result lists, so each covers the other's blind spot.
I
- Inference
- Inference is running a trained model to produce an output from an input, as opposed to training it, and it is what every API call or local run does.
J
- JSON Schema
- JSON Schema is a standard way to describe the shape of JSON data, its fields, types and required values, so it can be checked automatically.
L
- Large language model (LLM)
- A large language model is a neural network trained on a great deal of text to predict the next token, which lets it write, summarise, translate and answer questions.
- Latency
- Latency is how long a user waits for a response; for a language model it is usually split into time to first token and total time.
- LLM as judge
- LLM as judge is using a language model to score another model's output against a rubric when a simple code check cannot decide.
- Local model
- A local model is a language model that runs on hardware you control, a laptop, a server or your own cloud machine, rather than behind someone else's API.
- LoRA (low-rank adaptation)
- LoRA is a way to fine-tune a model by training small extra matrices beside its frozen weights, which needs far less memory than retraining everything.
M
- Model Context Protocol (MCP)
- The Model Context Protocol is an open standard for connecting AI applications to tools and data, so one server can serve any assistant that speaks it.
- Model drift
- Model drift is a change in how well an AI feature works over time, because the inputs, the world or the model behind an API has changed.
- Model routing
- Model routing sends each request to the model best suited to it, often a small cheap model for easy requests and a larger one only when needed.
- Multi-agent system
- A multi-agent system splits a task between several AI agents, each with its own role, tools and context, coordinated by code or by another agent.
O
- Observability
- Observability for AI features is recording each model call, its prompt, settings, tokens, time and outcome, so you can see what happened and why.
- Open weights
- An open-weights model is one whose trained parameters are published for download, so you can run, inspect and fine-tune it on your own machines.
P
- PII redaction
- PII redaction finds personal data, such as names, emails, phone numbers and addresses, in text and removes or masks it before the text is sent or stored.
- Prompt
- A prompt is the text you send a language model to tell it what to do: the instruction, any examples, and the material it should work on.
- Prompt injection
- Prompt injection is an attack in which text supplied by a user or found in data contains instructions that make a model ignore its own rules.
Q
- Quantisation
- Quantisation stores a model's weights with fewer bits, such as 4 instead of 16, so it needs less memory and runs faster with a small loss of quality.
R
- Rate limit
- A rate limit caps how many requests or tokens you may send to an API in a period of time; past it, requests are refused until the window resets.
- Red teaming
- Red teaming is deliberately attacking your own AI system, trying to make it misbehave, leak data or misuse its tools, so you find weaknesses before others do.
- Reranking
- Reranking takes the first results of a fast search and reorders them with a slower, more careful model so the best passages come first.
- Retrieval-augmented generation (RAG)
- Retrieval-augmented generation is answering a question by first finding relevant passages in your own documents and then giving them to a language model to answer from.
S
- Sandboxing
- Sandboxing runs a model's tools or generated code in a confined space, with limited files, network and time, so a mistake or an attack cannot spread.
- Streaming
- Streaming sends a model's reply piece by piece as it is generated, so a user sees the first words in a moment instead of waiting for the whole answer.
- Structured output
- Structured output is asking a language model to reply in a fixed machine-readable shape, such as JSON with named fields, instead of free text.
- System prompt
- A system prompt is the standing instruction sent before the conversation that sets a model's role, tone and rules for every reply in a chat.
T
- Temperature
- Temperature is a setting that controls how much randomness a model uses when it picks the next token: low is steadier, high is more varied.
- Text to SQL
- Text to SQL is a language model turning a question in plain words into a database query, which is checked before anyone trusts the answer it gives.
- Token
- A token is the piece of text a language model reads and writes in: often part of a word, a whole short word or a punctuation mark.
- Tool calling
- Tool calling lets a language model ask your program to run a named function with arguments, such as a search or a lookup, and then use the result.
- Transformer
- A transformer is the neural network design most language models are built on: a stack of identical blocks in which every token uses attention to take in what it needs from the others.
V
- Vector search
- Vector search finds the stored items whose embeddings are closest to the embedding of a query, so it matches by meaning rather than by exact words.
Comparisons
The choices people search for, answered in one sentence and a short table.
RAG vs fine-tuning
Use RAG to give a model facts it must look up and keep current; use fine-tuning to change how it behaves, such as a format, tone or narrow task.
MCP vs tool calling
Tool calling is how a model asks for a function to be run; MCP is a standard for offering such tools from a separate server that many AI applications can share.
AI agent vs workflow
A workflow runs steps you wrote in advance, using a model at some of them; an agent lets the model choose the steps, which is more flexible and harder to predict.
Local model vs API
A local model keeps data and versions under your control at the cost of hardware and operations; a hosted API gives larger models with no servers, charged per token.
Prompt engineering vs fine-tuning
Start with prompt engineering, which is fast and cheap to change; move to fine-tuning only when a measured gap remains that better prompts and examples cannot close.
Embeddings vs keyword search
Embedding search matches meaning, so it finds paraphrases; keyword search matches exact words, so it finds codes and names. Many systems use both together.
Evals vs unit tests
Unit tests check that code does exactly what it should every time; evals score how well a model-backed feature does over many cases, because its outputs vary.
Chatbot vs AI agent
A chatbot replies to each message in a conversation; an agent works towards a goal over many steps, choosing and calling tools, often with no message in between.
Open-weights vs closed models
Open-weights models can be downloaded, run and adapted on your own hardware within their licence; closed models are used only through the provider's API.
Streaming vs not streaming
Stream when a person is watching the reply appear, so they can start reading at once; wait for the whole reply when code must parse or check it first.