Glossary

Agent

A system that pursues a goal through state, action selection, observations, and stopping conditions. See AI Agents.

Attention

A mechanism that computes context-dependent combinations of value vectors using weights derived from query-key similarity. See Transformers.

Chunk

A retrievable unit created from a larger document. Chunk boundaries affect both retrieval precision and the amount of surrounding context preserved.

Context Window

The tokenized information a model can process for one generation, including input and reserved output. See Context Windows.

Embedding

A vector representation optimized so a chosen similarity function reflects useful relationships for a task. See Embeddings.

Grounding

Connecting generated claims to inspectable evidence or verified external state. See Grounding and Hallucination.

Hallucination

Fluent output that is false, unsupported, or misleading in its use context.

Large Language Model (LLM)

A neural model trained to estimate distributions over token sequences. See Large Language Models.

Prompt

The instructions, data, examples, and output requirements supplied to a model for a generation. See Prompting.

RAG

Retrieval-augmented generation: retrieving external evidence and including a selected subset in the generation context. See Retrieval-Augmented Generation.

Reranker

A model or scoring stage that reorders a limited candidate set produced by an initial retriever. See Reranking.

Structured Output

Model output constrained to a machine-readable contract such as JSON Schema or a typed record. See Structured Output.

Token

A discrete identifier produced by a tokenizer and processed by a model. It may represent a word, subword, symbol, whitespace, or bytes. See Tokenization.

Tool Use

A controlled interface through which a model-backed application requests an external function or service. See Tool Use and Function Calling.

Vector Database

A database or search system that stores embeddings and supports nearest- neighbor retrieval. See Vector Databases.