SIFS
Extremely fast hybrid code search for agents.
Start new thread
Tristan Manchester

SIFS - Extremely fast hybrid code search for agents.

by
Gravity Notes For Mac
•
Coding agents waste too much context before they understand a repo. They grep around the codebase, read large chunks, and make guesses. SIFS gives them fast local search. It runs as a CLI or MCP server. The default mode is hybrid search: BM25/semantic retrieval. Current benchmark: 63 large repos 19 languages 1,251 annotated search tasks index: 6.5 ms search: 0.376 ms SIFS is ahead of any comparable tool on this benchmark. Very open to feedback and issues/PRs! Let me know if you try it.

Add a comment

Replies

Best
Tristan Manchester
Maker
📌
I’ve been building SIFS, a Rust CLI/library/MCP server for local code search. The use case is coding agents. Before an agent reads half a repository, I want it to be able to ask rough questions like “where is authentication handled?” or “how does upload backpressure work?” and get ranked code chunks back. SIFS supports BM25, semantic search, and a hybrid mode. BM25 is fully offline and model-free. Semantic/hybrid search run locally once the model is cached. I’d especially like feedback on: - ranking failures - benchmark methodology - Rust API ergonomics - MCP interface design