Glossary¶

Vector store¶: A database that stores text chunks alongside their embeddings (numerical vectors). Queries are matched by semantic similarity rather than keyword search.
Embedding¶: A numerical representation of a piece of text. Embeddings capture meaning – similar texts have similar embeddings.
Chunk¶: A short section of a document (typically 200–500 tokens). Documents are split into chunks before embedding so that retrieval returns precise, relevant passages.
Collection¶: A named group of chunks within a vector store. One store can hold multiple collections (e.g. “nml-docs”, “nml-elife”).
Domain¶: A knowledge area with its own vector stores, description, and optional MCP tool configuration. Queries are routed to a domain by the classifier.
Guard model¶: A small LLM that screens queries for safety before they reach the main pipeline.
Chat model¶: The main LLM that classifies queries, generates search queries, and writes answers.
Evaluator¶: An LLM call that judges the quality of a generated answer and decides whether to accept it, retrieve more context, or regenerate.
MCP server¶: A Model Context Protocol (MCP) server that exposes tools the LLM can call live – e.g. querying a database, validating a file, or fetching data from an API.