Glossary

Key terms used throughout hypar and in RAG systems generally.

Term	Meaning
Chunk	A bounded piece of a document (~400 tokens). The unit of retrieval. See RAG Pipeline.
Embedding	A fixed-length vector (768 dimensions in hypar) that represents a chunk's semantic meaning. Produced by an embedding model.
pgvector	PostgreSQL extension that stores vectors and computes cosine distance. hypar's default vector store.
BM25	Classical keyword-relevance score; complements semantic search on names, codes, and literals that embeddings often miss.
Hybrid search	Weighted blend of vector similarity and BM25, controlled by the parameter alpha in [0, 1]. See Hybrid Search & HyDE.
HyDE	Hypothetical Document Embeddings: the LLM drafts a short hypothetical answer first, then embeds that text as the query vector. Often improves recall on factual Q&A.
MMR	Maximal Marginal Relevance: re-ranks retrieved chunks to balance relevance vs. diversity, controlled by lambda.
top-K	How many chunks survive filtering and are shown to the LLM as context.
Tool call	The LLM decides to invoke a function (here searchKnowledgeBase) instead of replying directly. Powered by the Vercel AI SDK.
Agent / search mode	auto = LLM decides whether to search; search = always retrieve from KB; direct = skip retrieval entirely. Configurable in Settings.
HNSW	Hierarchical Navigable Small World: the index type used on the embedding column for fast approximate nearest-neighbour search.
Cosine similarity	Similarity metric between two vectors. 1.0 = identical direction, 0.0 = orthogonal. hypar uses a score floor of 0.2.
Ingestion	The process of parsing a document into chunks, embedding each chunk, and storing them in the database. Runs as a durable background workflow.
Workspace	An isolated scope for documents and settings, shared among members with different roles. See Workspaces.