A plain-language, alphabetized glossary of AI and LLM terms including RAG, tokens, embeddings, context windows, temperature, and function calling.
Worked example: Example: look up context window before deciding whether a long document can fit in one model request.
An AI system that can plan and take a sequence of actions toward a goal, often using tools such as search, code execution, or external APIs. A useful agent has clear boundaries, observable steps, and approval points for consequential actions.
Link to this termThe maximum amount of input and output a model can consider in one request, measured in tokens. A larger context window can hold longer documents, but it does not guarantee the model will use every detail well.
Link to this termNumeric representations of text, images, or other data that place similar meanings close together in vector space. They are commonly used to retrieve relevant documents for RAG systems.
Link to this termFurther training a base model on a curated dataset so it follows a particular pattern or performs a narrower task more reliably. It is different from adding facts at request time through retrieval.
Link to this termA structured way for a model to request that an application run a named tool with typed arguments. The application, not the model, decides whether and how the requested action is executed.
Link to this termA model output that sounds plausible but is unsupported, incorrect, or invented. Retrieval, structured outputs, verification, and calibrated prompts can reduce but not eliminate this risk.
Link to this termThe process of running a trained model to generate an output from an input. API token pricing and rate limits usually apply during inference.
Link to this termA declarative format for describing the expected structure of JSON data, including object properties, types, and required fields. It is often used to constrain structured model outputs and tool arguments.
Link to this termThe instructions and context sent to a model for a single request. Good prompts state the task, relevant evidence, constraints, and desired output format.
Link to this termRetrieval-augmented generation: a pattern that retrieves relevant external information and includes it in a model request. It can ground answers in current or private documents without retraining the model.
Link to this termA model response constrained to a machine-readable shape such as JSON. It helps downstream software handle responses reliably, but the result still needs validation.
Link to this termHigh-priority instructions supplied by an application to guide a model’s role, behavior, or boundaries. It is part of the request context and consumes tokens.
Link to this termA sampling setting that influences how varied a model’s generated tokens can be. Lower values tend to be more consistent; higher values can increase variety and unpredictability.
Link to this termA unit of text processed by a model, often smaller than a word and sometimes spanning punctuation or whitespace. Providers typically price inputs and outputs by token count.
Link to this termA sampling setting that limits choices to the smallest group of likely next tokens whose combined probability reaches p. It is another way to balance consistency and variety.
Link to this termA storage system designed to index and search embedding vectors efficiently. It is often one component of a RAG pipeline, alongside chunking, metadata, and reranking.
Link to this term