DocuSense AI is a retrieval-augmented document intelligence system — parsing, embeddings, semantic search and extractive question-answering — that runs entirely client-side. Nothing you upload ever leaves this browser tab.
Load the bundled sample documents or drop in your own PDF/TXT/MD files. Every answer below is generated by models running in this tab right now.
Confidence: high >60% medium 30–60% low <30%
Drop files or click to upload
PDF, TXT, or MD · up to 10 MB per file
This simulates row-level permissions: documents tagged Restricted are excluded from retrieval unless the active role is allowed to see them — enforced at the retrieval step, not just in the UI.
Sources for your most recent question will appear here, ranked by cosine similarity.
No server round-trip. Every step below executes on-device using WebAssembly (ONNX Runtime Web) via Transformers.js.
If retrieval confidence is low, the system says so instead of guessing. Extractive answers can't hallucinate facts not present in your documents.
Every answer reports real latency and a real confidence score from the model — not a canned percentage.
Restricted documents are filtered out of the candidate set before search runs, not after — the model never even sees them.
A fixed set of 6 golden question/document pairs with known expected answers, scored automatically against whatever is currently in the document library.
This project follows a specific evaluation rubric: real architecture, real evaluation with baselines, honest error handling, and a working deployed demo — rather than a cosmetic chatbot UI.
Xenova/all-MiniLM-L6-v2 for embeddings and Xenova/distilbert-base-cased-distilled-squad for extractive QA, both quantized ONNX builds served via Transformers.js.
Retrieval Recall@1/@5, mean reciprocal rank, QA confidence score, and end-to-end latency — shown live in the Evaluation tab above, not claimed in prose.
Static HTML/CSS/JS. No database, no server, no secrets to rotate. Deployed to GitHub Pages and Vercel directly from this repository.