The watcher is what makes it worth wiring in permanently: edit a file and the next search sees the new version, so the index does not go stale between sessions. Results come back with the source path attached, which is what lets the assistant quote and then go read the rest.
A server that watches directories, embeds file contents into a vector store, and answers semantic queries over them.
- search — semantic search across the indexed files, returning the matched text, the source path, the file type and a score
- get_stats — how many documents are indexed, which directories are watched, and what is still processing
- Files re-indexed as they change, and removed from the index when deleted
- Chunking you control: CHUNK_SIZE defaults to 1000 and CHUNK_OVERLAP to 200
- IGNORE_FILE takes a .gitignore-style file to keep directories out of the index
Npx and the @lishenxydlgzs/simple-files-vectorstore package. You must set one of WATCH_DIRECTORIES, a comma-separated list of paths, or WATCH_CONFIG_FILE, pointing at a JSON file with a watchList array — the server has nothing to index otherwise. No account or key. search takes a limit that defaults to 5 and caps at 20.
One command — npx -y @lishenxydlgzs/simple-files-vectorstore