Glossary

Terms as the session uses them, in the order they appear.

Language model (LLM). A model that produces text given text. On its own it answers from what it absorbed in training, with no access to current data and no sources.

AI application. A system in which the model is one component. The other components select what the model sees, give it tools, check what it produces, and measure the result.

Instructions. The system message: what the application wants the model to do, including how to cite, what to do when information is missing, and when to ask.

Context. Text supplied to the model with the question: documents, passages, data. Selecting it is the job of the application.

Context window. The maximum amount of text a model can receive in one call. Cost grows with every token supplied.

Token. The unit in which models read and bill text; about four characters of English.

Metadata. Information about a document or a record: id, title, type, country, date, source, section. Part of what the application knows; used for filtering, ranking, and citations.

Chunk. The unit of text that retrieval returns and the model reads. Section chunks keep headings and meaning; fixed-size windows cut through both.

Embedding. A vector that places a text so that texts with similar meaning are close. Computed by an embedding model; TF-IDF is a word-based substitute.

Index. The chunks, their metadata, and their embeddings, held for search. In LangChain, a vector store such as InMemoryVectorStore, Chroma, FAISS, or pgvector.

Retrieval. Selecting the chunks that best match a question, by similarity of embeddings, with optional metadata filters and ranking adjustments.

Retrieval-augmented generation (RAG). Retrieve passages, supply them as context, generate an answer from them with citations.

Grounding. The property that an answer is written from the supplied evidence and cites it.

Tool. A function the application exposes to the model, with a description and a parameter schema. The model asks for a call; the application runs it and returns the result.

Tool call. The model's request to run a tool with given arguments. The application decides whether and how to run it.

Structured output. An answer in a fixed shape, defined by a schema and validated, so that other software can read it.

Schema. The definition of a shape: fields, types, required values. Pydantic models produce JSON schemas.

Model Context Protocol (MCP). A standard for exposing tools, resources, and prompts to AI clients, so that one server serves many clients. The session's indicator server is written with FastMCP; the Data360 MCP server is the program's deployed example.

Workflow. A sequence of calls written in code, with the model used at fixed points. Predictable and testable.

Agent. A loop in which the model chooses which tools to call and when to stop. Adapts to the question; needs a step limit, a trace, and tools that return errors as data. Built in the session with LangGraph through create_agent, bounded by recursion_limit.

Trace. The record of every tool call and result in a run, kept so that a reviewer can see what the application did.

Claim. A number in an answer tied to the record it came from, by an identifier such as IT.NET.USER.ZS:KEN:2024.

Verification. Comparing each claim with its record under a policy (rounded to a number of decimals, or exact). Verified: matches. Flagged: differs. Unmarked: no claim, so no mark.

Fail-closed. The rule that a number without a claim receives no mark, whether or not it happens to be right.

Proof-Carrying Numbers (PCN). The program's protocol and packages for claims and verification in generated text.

Evaluation suite. Questions with expected evidence, expected numbers, and expected behaviour, run against the application after every change.

Retrieval hit. For one question, whether the top-k passages include a passage from an expected document.

Threshold. The retrieval score below which the application treats the result as no evidence.