The vocabulary of local hybrid search.

Hybrid search

A retrieval strategy that combines keyword (lexical) matching with semantic (vector) similarity, and optionally metadata signals, into a single ranked result set. StrataFS fuses FTS5 BM25, vector cosine similarity, and metadata scoring in one SQL query so exact matches and meaning-based matches surface together.

BM25

Best Matching 25 — the ranking function behind classic full-text search. It scores documents by term frequency and inverse document frequency, favouring rare, on-topic words. StrataFS gets BM25 from SQLite's FTS5 extension, so exact keyword matches stay fast and precise.

Vector search

Retrieval by semantic similarity: text is turned into a high-dimensional embedding, and results are ranked by how close their vectors are (usually cosine similarity) to the query's vector. It finds relevant passages even when they share no keywords with the query.

Embedding

A dense numeric vector that represents the meaning of a piece of text. StrataFS generates embeddings locally with FastEmbed + ONNX — no external API — and stores them in SQLite via sqlite-vec for cosine-similarity ranking.

sqlite-vec

A SQLite extension that stores vectors and runs similarity queries inside the database. It lets StrataFS keep embeddings alongside full-text and metadata in one file, so there is no separate vector database to deploy or operate.

FTS5

SQLite's full-text search module. It provides tokenised indexing and BM25 ranking for keyword queries. StrataFS pairs FTS5 with sqlite-vec so lexical and semantic search share the same embedded engine.

Chunking

Splitting a document into smaller passages before indexing, so search can return the most relevant section rather than a whole file. StrataFS uses streaming chunkers that keep memory flat even on large files.

Reranking / score fusion

Combining several ranking signals into one ordered list. StrataFS fuses BM25, vector cosine, and metadata scores with tunable weights, so you can bias toward exact matches, semantic matches, or recency per query.

Metadata filter

Constraining or scoring results by structured attributes — file type, source, path, or modified date — rather than content alone. In StrataFS, metadata is a first-class ranking signal fused into the hybrid query, not a bolt-on post-filter.

MCP (Model Context Protocol)

An open protocol for letting AI agents discover and call tools over JSON-RPC. StrataFS ships a native MCP server exposing tools like stratafs.search, so agents such as Claude Desktop or Cursor can query your files directly.

RAG (Retrieval-Augmented Generation)

A pattern where a model retrieves relevant documents and generates an answer grounded in them, instead of relying only on training data. StrataFS is the local, offline retrieval layer of a RAG stack, feeding grounded context to an LLM.

Semantic filesystem

A filesystem you query by meaning instead of by exact path or name. StrataFS indexes your files' contents and exposes them through search and an MCP server, so you (or an agent) can ask a question and get the right file back.

See it in practice: the search engine, the architecture, or the quickstart.