Summary
Five hundred documents you can ask questions of, with nothing leaving the machine.
It indexes what is already on your disk and returns the relevant passages, so the assistant answers from your contracts and reports instead of from its training data. After the one-time model download there are no network calls at all.
What it is
A local semantic index over a folder of documents — txt, md, pdf, docx, pptx, csv — stored as LanceDB files on disk and searched in natural language.
What you get
- A folder indexed or re-indexed, with the supported document types picked up automatically
- Natural-language search that returns the relevant chunks, not whole files
- Hybrid search: semantic similarity combined with full-text keyword search via Reciprocal Rank Fusion
- Incremental indexing by content hash — only changed files are reprocessed
- Index status: total chunks and which files are in it, with their chunk counts
- A single file forced back through indexing, ignoring the hash cache
- 50+ languages, including asking in one language and finding results in another
- Fully offline after a one-time model download
Requirements
A recent Python, and room on disk for the model and the index. LanceDB keeps everything as local files, so there is no server, container or database to run, and no API key.