This semantic search tool lets you ask questions of your own documents privately. Drop a folder of PDFs, Word files, notes, HTML or CSV, and search by meaning — “how many holidays do I get” finds “20 days of paid annual leave” — using a small embedding model that runs in your browser. Keyword search works instantly with no download. Results show the file, the passage and its score, and export as CSV.
| Query | Keyword finds | Meaning also finds |
|---|---|---|
| holidays | “holidays” | “annual leave” |
| refund | “refund” | “money back”, “reimbursed” |
| password rules | “password” | “14 characters”, “two-factor” |
Search that matches meaning rather than exact words, using vectors that place related phrases close together.
PDF, Word (.docx), text, Markdown, HTML, CSV, JSON, logs and subtitles.
No, the index lives in this browser tab and disappears when you close it.
No. Files and text are processed in your browser. Only the model itself is downloaded, once, when you ask for it.
Yes. The Downloaded models panel lists what this tool stored and deletes it after a warning; you'd need to download it again to use it.
Files are read locally (PDF.js for PDFs, mammoth for Word, DOM parsing for HTML, plain text otherwise) and split into overlapping passages of about 600 characters. Keyword mode ranks passages with BM25 (k1 = 1.2, b = 0.75) over lower-cased, accent-stripped, lightly stemmed tokens. Meaning mode embeds every passage and the query with a sentence-transformer (all-MiniLM-L6-v2, or a multilingual MiniLM) through Transformers.js with mean pooling and normalisation, then ranks by cosine similarity blended 80/20 with the keyword score. The index exists only in the open tab.