AI Engineering Glossary

Plain-language definitions of the terms that come up most often in AI engineering discussions.

Short, plain-language definitions of the terms that come up most often in our workshops and community discussions. When a term here has a full workshop of its own, the entry links to it.

Core terms

  • LLM (large language model) — a neural network trained on large amounts of text to predict the next token. Chat assistants, coding helpers and most AI features are built on top of one.
  • Token — the unit of text an LLM reads and produces. Roughly three quarters of a word in English; APIs price and limit context in tokens.
  • Prompt — the text you send to a model. Prompt engineering is the discipline of structuring prompts so outputs stay useful and predictable.
  • Context window — how much text a model can consider at once. Everything the model "knows" about your request must fit in it.
  • RAG (retrieval-augmented generation) — fetching relevant documents and putting them into the prompt before the model answers, so answers are grounded in your data instead of the model's training set.
  • Embedding — a vector that represents the meaning of a piece of text, so similar texts land close together in vector space.
  • Vector search — finding the nearest embeddings to a query embedding. The retrieval half of most RAG systems.
  • Agent — an LLM wrapped in a loop that can call tools (search, code execution, APIs) and decide the next step itself until a goal is met.
  • Function calling — the structured way a model asks your code to run a tool: it outputs JSON matching a schema you declared, your code executes it and returns the result.
  • Guardrails — checks around a model's input and output that block unsafe or off-spec results before they reach users.
  • Evaluation (eval) — an automated test for model quality: a dataset of cases, a scoring function, and a threshold that fails the build when quality drops.
  • Fine-tuning — continuing to train a base model on your own examples so it adopts a style or format, cheaper than retraining but narrower than prompting for knowledge.

Shipping terms

  • Latency — how long one model call takes. Drives whether a feature feels instant or batch.
  • Hallucination — a fluent but wrong answer. Mitigated with grounding (RAG), guardrails and evals — never fully eliminated.
  • Human in the loop — a design where a person reviews or approves model output before it takes effect.