On-device AI · WebGPU

Search your PDFs right in your browser.

Client-side PDF RAG. Extract, embed, and retrieve — all on-device, no servers.

WebGPU / WASM
MiniLM embeddings
100% private
Why RAG Search

Semantic search. Zero uploads.

Upload PDFs, ask questions, and get answers with citations — all locally.

100% private

Your documents never leave your device. All processing happens locally.

On-device embeddings

Text embeddings generated locally via Transformers.js. No API calls.

Chunk & retrieve

Automatic text chunking and vector similarity search for precise retrieval.

Cited answers

Answers include source citations so you can verify the context.

How it works

Three steps. Zero servers.

  1. 01

    Upload a PDF

    Drop in any text-based PDF. Text is extracted page by page in your browser via pdf.js.

  2. 02

    AI builds the index

    Text is chunked and embedded on-device with all-MiniLM-L6-v2 via WebGPU, stored in an in-memory vector store.

  3. 03

    Ask & retrieve

    Type a question. The top matching passages are ranked by cosine similarity and shown instantly.

Complete guide

About document RAG search

A free “chat with your PDF” tool that builds a semantic search index over your documents entirely in the browser. Instead of keyword matching, it finds passages by meaning — ideal for research papers, contracts, manuals, and long reports where the answer exists but you do not know the exact words to search for.

How it works

pdf.js extracts the text layer of your PDF on-device. The document is split into chunks and embedded with all-MiniLM-L6-v2, a compact sentence transformer running via Transformers.js with WebGPU acceleration; the embeddings live in an in-memory vector store. When you ask a question, your query is embedded with the same model and ranked against every chunk by cosine similarity, returning the top matching passages with page numbers and scores in seconds.

Limits & requirements

The tool reads PDFs with a selectable text layer — scanned, image-only PDFs need OCR first. There is no hard file-size limit, but very large documents scale with your device’s available RAM. Chrome or Edge with WebGPU is recommended for fast indexing; other browsers work via WebAssembly.

Privacy

The model downloads once from the open-source model hub, then indexing and search are 100% local. Confidential documents never touch a server, and nothing about their content is collected.

Support

Questions, answered.

No. Everything runs in your browser. The embedding model downloads once from Hugging Face and is cached locally — after that, indexing and search are 100% on-device.