DocChat
Chat with your PDFs in the browser: answers with clickable page citations, plus summaries, quizzes and PDF comparison. No upload, account or API key.
Still loading. If the tool doesn’t appear, reload this page.
About this tool
DocChat
Ask questions about your PDFs and get answers with page-level citations. A language model, search and PDF reader all run on your device: no upload, no account, no API key.
DocChat is a private PDF workspace where you read a document and talk to it at the same time. Add one or several PDFs, ask a question in plain language, and the answer comes back with numbered citations. Click a citation and the PDF jumps to that page with the supporting sentence highlighted. The same tool writes full-document and chapter-wise summaries, lists key points, finds and defines terms, searches inside your files, compares two or more PDFs, and turns a document into a multiple-choice quiz. Everything runs on your device: PDF.js reads and draws the file, a small embedding model finds the passages that matter, and a language model running on your graphics card (WebGPU) writes the answer from those passages only. If the document does not contain the answer, DocChat says so instead of guessing. Models download once and are cached by the browser, so with the page open they then work without a connection. Your documents, questions and answers never leave this page, and you can delete them at any time. Its AI models download from huggingface.co, and nothing you add or ask is sent there. The language model's compiled WebGPU code downloads from raw.githubusercontent.com.
How to use DocChat
- 1
Drop one or more PDFs onto the page or choose Browse files. DocChat reads the text, splits it into passages and builds a private index on your device.
- 2
Ask a question, or use Summarize, Chapter summaries, Key points, Quiz or Compare. The first time, load a language model from the model panel; until then you get the matching passages.
- 3
Click any citation to jump to the page with the evidence highlighted, or search inside your documents from the Search tab.
- 4
Copy an answer, download a summary, export the conversation, or delete your documents and models when you are done.
Why use DocChat
Every answer is checkable
Each statement carries a numbered citation. Clicking it opens the PDF at that page and highlights the sentence that supports it. Citations that do not match a retrieved passage are dropped, and when nothing relevant is found the answer says so.
Private by architecture
Text extraction, search indexing, embeddings and the language model all run in your browser. There is no server to send your files to, no account and no API key. Removing a document deletes its text, index and any answers that quote it.
Retrieval you can trust
Passages are found by meaning (local embeddings) and by keywords (BM25), fused into one ranking. A passage is only used when at least one signal says it is genuinely related to your question.
Works with several PDFs
Ask across all documents or just the one you are viewing. Comparison pulls passages from each file separately so facts are not mixed up, and every citation names its file.
Free, then offline
The embedding model (23 MB) and an answer-writing model (from about 300 MB) download once from Hugging Face and are cached. After that, with the page open, DocChat answers without an internet connection.
Common use cases
- You have a contract, report, paper or manual and need an answer with proof of where it came from, without uploading a confidential file to a chatbot.
- You want a quick summary, the key points or a chapter-by-chapter outline of a long PDF, with page numbers to check.
- You are comparing two versions or two documents and want the similarities and differences, each tied to a page in the right file.
- You study from PDFs and want a quiz generated from the text, with a link back to the page that backs every answer.
Frequently asked questions
Are my PDFs uploaded anywhere?
No. The PDF is read, indexed and answered from inside your browser. The only network traffic is the one-time download of the model files from Hugging Face and the model runtime from GitHub, and neither of those receives your documents.
Why does a model have to be downloaded?
The language model is what writes answers, and it runs on your own graphics card, so its file must be on your device. Qwen2.5 0.5B is about 290 MB and the recommended 1.5B is about 880 MB. They are stored in the browser cache and reused offline.
What if my browser has no WebGPU?
DocChat still reads, searches and highlights your PDFs. Instead of a written answer you get the passages that best match your question, with page numbers and highlighting. Summaries use sentences copied from the document. Use a recent desktop Chrome or Edge for written answers.
How do I know the answer is not made up?
Answers may only use the numbered passages supplied to the model, and each claim should cite one. A citation chip opens the exact page and sentence. If a statement has no citation, DocChat warns you, and small models can still make mistakes, so check the highlighted text.
Can it read scanned PDFs?
Only PDFs that contain text. A scan made of images has nothing to extract, and DocChat does not run OCR. It tells you when pages have no selectable text.
How big a PDF can I use?
Up to 120 MB and 2,500 pages each, with up to 20 documents at once. Large files take longer to index because every page is read and every passage embedded on your device.
Does it work offline?
Yes, once the models have been downloaded and while the page is open: load DocChat while you are online, then add or open documents and ask questions without a connection. Opening or reloading the page itself still needs the site.
The model download fails or says it cannot be stored. What can I do?
Browsers limit how large a cached file may be by the free space left on your disk, so a model cannot be stored when the disk is nearly full. DocChat first tries the browser cache, then falls back to the origin private file system and IndexedDB, which have no such limit. If every option fails, free a few gigabytes of disk space and try again, or choose the smaller Qwen2.5 0.5B model.
Can I try DocChat without a PDF?
Yes. Choose Try a sample on the first screen: DocChat builds a short made-up report as a real PDF in your browser and opens it, so you can ask questions, click citations and see the highlighting straight away.
How do I choose and download an AI model?
Click Get AI answers in the header. The model centre checks your device, recommends a model, and shows Fast, Balanced and Best choices with their download size and how much speed, quality and memory each needs. Pick one and press Download: you see MB fetched, speed and time left, you can close the window and keep working while it downloads, and you can cancel at any time.
How do I delete my data?
Use the bin next to a document to remove it with its text, index and answers that quote it, or Manage data to delete everything including downloaded models.