A private ChatGPTfor any document.
Drop in a .pdf, .txt, or .md and ask questions about it. Semantic retrieval and the language model both run in your browser, so the answer is grounded, cited, and the file never leaves the tab.
- 0Server calls
- 2On-device models
- WebGPULocal LLM
Drop a PDF or text file here, or click to choose
.pdf · .txt · .md — stays on your device — up to 15MB
First use downloads the embedding model (~23MB), then an answer model (~250MB) on your first question — once, then cached. No WebGPU detected — answers can still run on your CPU with the lighter SmolLM2-360M (slower), offered after the first search.
Not search. Retrieval-augmented generation.
- 01
Nothing is uploaded
The file is read, chunked, embedded, retrieved, and answered locally. NotebookLM and ChatGPT can do document Q&A too — but only after you hand them your file.
- 02
Meaning, not Ctrl+F
Ask "how do I cancel" and it finds the passage about "terminating your subscription" — hybrid vector + BM25 retrieval, fused with Reciprocal Rank Fusion.
- 03
The model runs here too
A small instruct model (Qwen2.5-0.5B) streams a grounded answer with citations over WebGPU. No API key, no round-trip, works offline once cached.