Inquire — ask any document
Paste a contract, report, or article, ask a question in plain English, and get the exact answer pulled straight out of the text and highlighted. A real reading-comprehension model does the thinking, entirely on your machine.
🔒 100% private — the document and your questions never leave your deviceWhy answer questions in the browser?
Every popular "chat with your document" tool uploads your file to someone else's servers. That is fine for a blog post, but not for a lease, an offer letter, a medical result, a legal brief, or an internal report. Inquire flips that around: the AI model itself is downloaded to your browser once, and then every question you ask is answered locally. Nothing is uploaded, there is no account, and it keeps working with your Wi-Fi off.
How it works
Inquire runs a DistilBERT model fine-tuned on the SQuAD reading-comprehension dataset. It does extractive question answering: instead of writing a fresh paragraph the way a chatbot would, it finds the span of text in your document that best answers the question and hands it back with a confidence score. That makes the answers traceable — you can see exactly where in the source they came from, highlighted in place.
The model, its tokenizer, and the ONNX runtime are loaded with Transformers.js and executed on WebAssembly inside a Web Worker, so a long document never freezes the page. To handle text that is longer than the model's context window, Inquire splits it into overlapping windows, scores every window against your question, and returns the single most confident answer across the whole document — so a detail buried on the last page is found just as easily as one in the first line.
Good questions to ask
It is strongest on factual, extractive questions: amounts, dates, names, deadlines, percentages, definitions, and "who / what / when / how much" style queries where the answer is actually written somewhere in the text. It is not a chatbot and will not invent facts or summarize an argument — if the answer is not in the document, it tells you its confidence is low rather than making something up.
Is my document uploaded anywhere?
No. The model runs locally in your browser with WebAssembly. Your document and your questions never touch a server, which is the whole point of building it this way.
How can an AI model run without a server?
The DistilBERT question-answering model is about 65 MB of ONNX weights. It downloads once, is cached by your browser, and then runs on WebAssembly in a background Web Worker. After the first load it works entirely offline.
Why does it highlight text instead of writing an answer?
This is extractive QA. It locates the exact passage that answers your question so the answer is verifiable and can never hallucinate a fact that is not in your document. You always see where the answer came from.
Does it handle long documents and PDFs?
Yes. Drop a .txt, .md, or text-based .pdf and it is read locally. Long text is split into overlapping windows so nothing at the end gets cut off, and the highest-confidence answer across the whole document wins.
Which browsers work?
Any recent desktop Chrome, Edge, Firefox, or Safari with WebAssembly and Web Workers. It runs on mobile too, though the one-time model download is heavier there.
I build a tool like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.