Private Document Q&A

What "Private Document Q&A" Means

Cloud-based "chat with your PDF" tools typically require you to upload the very file you want to keep confidential — contracts, financial statements, medical letters, internal reports. This tool flips that threat model. The document is parsed and indexed in your browser, the embedding model and the language model are downloaded once and cached in IndexedDB, and every question is answered without a single byte leaving the tab. No account. No telemetry. No third-party API call.

The On-Device Models

How to Use Private Document Q&A

1. Drop a .pdf, .docx, .txt, or .md file into the upload area (up to 30 MB).
2. Wait for the indexer — the document is split into sentence-aware passages and each one is embedded into a 384-dimensional vector on your device.
3. Type a question about the document. The tool retrieves the three most relevant passages and passes them to the answer model.
4. Read the grounded answer. Expand Show source passages to verify every claim against the original text, with page numbers for PDFs.
5. Ask follow-up questions — the document index is reused, so only the answer step runs each time.

Why Retrieval-Augmented Generation Beats a Plain LLM

Generative language models are notoriously prone to hallucination — they confidently invent facts when the prompt is ambiguous. Retrieval-augmented generation (RAG) sidesteps that by attaching the most relevant document excerpts to the prompt itself. The language model is no longer "answering from memory"; it is summarising the excerpts you can see for yourself. The system prompt also explicitly instructs Llama 3.2 to refuse when the answer is not in the excerpts — so unanswerable questions return "I could not find this in the document." rather than a fabricated reply.

Key Features

• Four formats in — accepts .pdf, .docx, .txt, and .md documents up to 30 MB.
• Grounded answers — every reply is generated from passages retrieved by cosine similarity against the document, not from the model's parametric memory.
• Inline citations — each answer expands to reveal the exact source passages, with page numbers for PDFs, so you can verify every claim.
• Offline after first load — the ~1 GB combined model bundle (MiniLM + Llama 3.2 1B) is cached in IndexedDB and works offline forever after.
• Reusable index — once a document is embedded, every follow-up question only re-runs the small embedding pass and the answer pass.
• Browser-only privacy — no upload, no account, no API key, no telemetry, no third-party service in the loop.

When to Use Private Document Q&A

• Reviewing a contract — ask "what is the notice period for termination?" without sending the contract to a hosted AI.
• Skimming research papers — ask "what dataset did the authors use?" or "what were the main findings?" on a 30-page PDF.
• Onboarding documents — ask an employee handbook or a long PRD targeted questions instead of scrolling.
• Compliance reviews — query an internal policy without copying snippets into ChatGPT or Claude.
• Financial statements — ask "what was the operating margin in Q3?" against a 10-Q PDF.

How It Compares to "Chat With Your PDF" Services

Hosted services like ChatPDF, Humata, and AskYourPDF upload your document to their servers, store it in a vector database they own, and run a third-party LLM (typically OpenAI or Anthropic) over the retrieved chunks. The privacy contract is "we promise not to look at it." This tool replaces every step of that pipeline with an in-browser equivalent: pdf.js parses the file, MiniLM embeds the chunks into IndexedDB-cached memory, cosine similarity runs over Float32 vectors in JavaScript, and Llama 3.2 1B generates the answer in a Web Worker. Pair it with the Document PII Redactor if you need to strip personal data from the file before storing the redacted copy elsewhere, or the PII Masker for AI Prompts when you do need to use a hosted LLM and want to scrub names, emails, and IDs first.

Limits and Trade-offs

• Document length. The index is capped at 80 passages (~96 000 characters of content). For very long PDFs split the file before uploading.
• Adaptive context. Documents under ~12 000 characters (~3 000 tokens) skip retrieval entirely — the full document is passed to Llama in document order. Longer documents fall back to top-3 cosine-similarity retrieval. The answer card labels which mode ran.
• Top-k window. When retrieval runs, only the three highest-similarity passages are passed to the answer model. Questions that need synthesis across many paragraphs may need to be asked more narrowly.
• Reasoning depth. Llama 3.2 1B is a small on-device model — it is strong at extracting and lightly synthesising facts from the supplied passages but not at deep multi-step reasoning. For complex chained logic, use the answer as a starting point and verify against the cited excerpts.
• First-time download. Around 1 GB total the first time (23 MB MiniLM + 990 MB Llama 3.2 1B in q4f16). Cached forever after, and re-used by the Resume Builder and Cover Letter Generator on this site.
• Browser support. Chrome, Edge, Firefox, and Brave are supported. Safari is hard-gated until JavaScriptCore reclaims WebAssembly memory during decoder runs.