Search your own PDFs with answers that cite sources.

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

PDF, DOCX, plain-text, and Markdown files are all supported, and several documents can share one searchable corpus. There is no account, no API key, and no server that ever sees your files.

How it works

Each document is parsed on-device: pdf.js walks the PDF text layer page by page, DOCX files are unpacked locally, and text and Markdown files are read directly. The text is cut into overlapping chunks of about 150 words (with a 30-word overlap so context is not lost at a boundary), and each chunk is embedded with Xenova/all-MiniLM-L6-v2, a compact 384-dimensional sentence transformer running through Transformers.js.

The embeddings live in an in-memory vector store in the tab. When you ask a question, your query is embedded with the same model and compared against every chunk by cosine similarity; the highest-scoring passages come back in seconds, each tagged with its source document, page number, and match percentage. Query terms are highlighted inside each passage, and embeddings are computed in batches of 16 chunks to keep long corpora responsive.

One corpus, many documents

You are not limited to a single PDF. Add a contract, its appendices, and your own notes on top, then search across all of them at once — the corpus panel lists each document with its page count, lets you remove a single source (its chunks disappear from the index immediately), or start a fresh corpus in one click. Stats for document count, pages, chunks, and vector dimensions stay visible as the corpus grows.

The number of results is yours to tune: choose the top 3 for a focused answer, 5 (the default) for everyday use, or 8–10 when you are surveying everything relevant across a large corpus.

WebGPU and the WASM fallback

Chrome and Edge with WebGPU embed on the GPU, which makes indexing a few-hundred-page document feel close to instant. Safari, Firefox, and older machines fall back to a WebAssembly build of the same model — slower, but feature-identical. The embedding model (roughly 25–90 MB of files depending on the backend) downloads once from the open model hub and is cached by the browser, after which indexing and search work entirely offline.

Scanned PDFs and other limits

Retrieval reads the text layer, so an image-only scanned PDF has nothing to index — the tool detects this and points you to the OCR tool, after which the recognized text can be added like any other document. There is no fixed file-size cap, but an in-memory index scales with your device’s RAM; if a very large corpus strains the browser, removing a document frees its vectors instantly.

Free and Pro limits

The free tier holds a corpus of 5 documents with up to 150 pages per document, with unlimited searches and no daily quota. Pro raises the corpus to 20 documents and 1,000 pages per document. A document past your page limit contributes its first pages and tells you plainly how many were skipped; both tiers include commercial use and carry no watermark.

Privacy

Confidential documents — NDAs, medical reports, unpublished research — never touch a server. After the one-time model download, indexing and search are 100% local, and closing the tab discards the in-memory index entirely. Nothing about document content is collected, logged, or used to train anything.