Loading page
About this tool
Ask local documents questions and receive deterministic matching excerpts with source citations—without a cloud model or upload.
The Private Document Chat builds a searchable index of your own documents inside the browser and answers a question by returning the exact passages that match it, each labelled with its source file and section — it retrieves, it never generates. Text is split into roughly 1,200-character chunks with 160 characters of overlap at sentence and line boundaries, then ranked with BM25 (k1 = 1.2, b = 0.75) plus bonuses for covering more of your query terms and for containing the query as a literal phrase. It suits anyone who wants to search a contract, a policy pack or research notes without uploading them to a cloud model.
Open Private Document Chat on AltFTool — it loads instantly in your browser.
Click "Choose up to 8 local documents" and pick TXT, Markdown, CSV, JSON, text-based PDF or DOCX files — 8 MB each and 20 MB in total; scanned image-only PDFs will not work because there is no OCR step.
Press "Build private index" to chunk and index the text inside the page, then type your question into "Ask with document-specific words" and press "Find grounded excerpts".
Each hit is a verbatim passage with its source document and section cited and a BM25 relevance score, with no answer generated; "Export counts only" saves a JSON report of document, chunk and search counts, and "Reset index" clears everything.
Each result names the source document and the section it came from, so you can verify the passage against the original.
The output is the document's own text, ranked — there is no model to hallucinate a clause that was never written.
BM25 scoring, term-coverage weighting and an exact-phrase bonus push the passage that actually answers the question to the top.
TXT, Markdown (.md/.markdown), CSV, JSON, text-based PDF and DOCX. Scanned or image-only PDFs will not work because there is no OCR step — the text has to already be selectable in the file.
Up to 8 documents, each 8 MB or smaller, with a 20 MB combined limit. The index is additionally capped at 250,000 characters per document, 500,000 characters in total, and 2,000 chunks.
It finds text. The tool returns up to 8 ranked source excerpts and does not summarise, infer, or verify them, which is why nothing it shows can be a fabricated answer. Reading and judging the passages is your part.
No. Extraction, indexing and retrieval all run as JavaScript in the page, with no cloud model involved. The optional session report is counts-only — it records how many documents, chunks and searches there were, and deliberately excludes file names, questions, excerpts and document text.