The AI glossary.

A model given a goal, a set of tools it can call, and a loop — it decides which action to take next, runs it, reads the result, and repeats until the task is done or it gives up. The step up from a single prompt-and-answer to something that acts.

The mechanism that lets a model weigh which earlier words matter most when predicting the next one. The T in the Transformer, and the reason modern models handle long, context-dependent text.

How much text a model can hold in mind at once, measured in tokens. Everything — your prompt, the documents you paste, the conversation so far, and the reply — has to fit inside it. Run over, and the earliest content falls out.

Prompting a model to work through a problem step by step before answering, instead of jumping to a conclusion. It trades speed and tokens for accuracy on anything that needs real reasoning.

A list of numbers that captures the meaning of a piece of text, so that similar meanings sit close together. The thing a vector database stores and searches — the foundation of semantic search and RAG.

A test suite for a model or an agent. You define what a good answer looks like, run the system against a fixed set of cases, and score the output — so a prompt or model change becomes a number you can compare, not a hunch.

Continuing to train a base model on your own examples so it picks up a style, format, or narrow skill. Powerful, but usually the last resort — prompting and retrieval solve most problems first, and cheaper.

Letting a model trigger real code — a search, a database query, an API — by returning a structured request instead of prose. The bridge between a language model and the systems it needs to actually do the job.

Shaping content so AI answer engines — ChatGPT, Gemini, Perplexity, Google's AI Overviews — cite it. The successor to SEO for a web where the model answers first and the click is rare.

Checks that sit around a model to keep its inputs and outputs in bounds — blocking unsafe requests, catching leaked data, forcing a valid format. What keeps a demo from becoming an incident in production.

When a model states something false with full confidence — an invented citation, a wrong number, a function that doesn't exist. It isn't lying; it's predicting plausible text. Retrieval and evals are how you catch it.

Running a trained model to get an answer — as opposed to training it in the first place. It's the part you pay for on every request, and where latency and cost live.

A cheap way to fine-tune: freeze the giant base model and train a small set of add-on weights instead. Fraction of the compute, and you can swap adapters in and out for different tasks.

An open standard for connecting models to tools and data through one common interface, so you don't hand-wire every integration. Think of it as a universal port for giving an agent new capabilities.

An attack where hidden instructions in content the model reads — a web page, a document, an email — hijack what it does next. The top security risk for anything that lets a model act on untrusted text.

Shrinking a model by storing its weights at lower precision — 16-bit down to 8 or 4. It cuts memory and cost, and lets bigger models run on smaller hardware, for a small hit to quality.

Fetching relevant documents at question time and handing them to the model as context, so the answer is grounded in your data instead of the model's memory. The standard cure for hallucination and stale knowledge.

A second-pass model that re-orders search results by how well each one actually answers the query. Bolted onto RAG, it lifts the best passages to the top before the model ever reads them.

The standing instructions a model reads before any user message — its role, rules, and tone. It sets the frame for the whole conversation, and it's where most of the behavior you notice is actually decided.

The unit a model reads and writes — a chunk of text roughly ¾ of a word. Context windows are measured in tokens, and API bills are counted in them, so it's the number that quietly governs both limits and cost.

A dial from predictable to creative. Low temperature makes a model pick the safest next token every time; raise it and the output gets more varied and surprising — useful for brainstorming, risky for facts.

A database built to store embeddings and find the closest ones fast. It's what makes semantic search and RAG work at scale — Pinecone, Qdrant, Weaviate and Chroma all live in this box.