AI Document Chat

About AI Document Chat

AI Document Chat is a private ChatGPT-style assistant for your own files. Drop in contracts, reports, manuals, meeting notes, or research papers and ask questions in plain English — "What's the termination clause?", "Summarize the findings", "What did we agree to pay?". The AI reads your documents, answers your question, and shows you the exact passages it relied on.

The difference from every other "chat with PDF" service: your documents never leave your computer. The language model itself — an open-source AI — downloads once into your browser's cache and runs right on your own device — on the graphics chip via WebGPU, or on the CPU with a compact model if your browser doesn't support WebGPU. There is no server, no upload, no account, and no per-question fee. After the first load it works completely offline: you can literally turn off your Wi-Fi and keep chatting.

What you can do

• Chat with sensitive files safely: contracts, financial statements, medical records, HR documents — nothing is transmitted anywhere, ever.
• Multiple documents at once: load several PDFs, Word docs, and text files and ask questions across all of them.
• Cited answers: every answer lists the document passages it used, with page numbers for PDFs — click to read the original text.
• Three model sizes: Fast (~0.4 GB) for quick answers on modest hardware, Balanced (~0.9 GB, recommended), and Best (~2.1 GB) for the highest quality on capable machines.
• Fully offline after first load: the model and search index are cached on your device — perfect for planes, secure environments, or unreliable connections.
• General assistant too: no documents loaded? It works as a normal private AI chat.
• Export your conversation as a Markdown file, or clear it instantly — chat history is never stored.

How it works

• Your documents are parsed in the browser (PDF.js for PDFs, mammoth.js for Word) and split into passages.
• A small embedding model indexes the passages so the most relevant ones can be found for each question — semantic search, on your device.
• An open-source language model (Qwen 2.5 or Phi 3.5) runs on your graphics chip via WebGPU — or on your CPU if WebGPU isn't available — and writes the answer from those passages, citing each one it used.

Requirements & tips

• Works on almost anything: with WebGPU (current Chrome, Edge, Firefox, or Safari) models run on your graphics chip. Without WebGPU the tool automatically switches to a compact model on your CPU — answers come slower, but everything still works and stays private.
• The first load downloads the model — do it on Wi-Fi. Every later visit starts instantly from cache.
• Start with the Balanced model. Step down to Fast if loading fails or answers are slow; step up to Best on a gaming-class machine for noticeably smarter answers.
• Scanned PDFs contain images, not text — run them through our PDF OCR tool first, then load the result here.
• Small on-device models are excellent at finding and summarizing what's in your documents; for deep general-knowledge reasoning, a larger cloud AI will still do better. The privacy trade is the point.
• Always verify critical details against the cited passages — that's what the citations are for.

No accounts, no upload queue, no per-document fees, no "free trial pages" — and no possibility of a data breach, because there's nothing to breach. Your documents stay where they belong: with you.

Related tools