How it works

Five steps, all on your machine. Vercel only serves static files; there are no functions, keys or costs.

Reads the PDF
pdf.js extracts the text page by page inside a Web Worker. The file is never sent anywhere.
Chunks and indexes it
The text is split into chunks of about 350 tokens that never cross a page boundary, and a multilingual embedding model turns each chunk into a vector.
Remembers it
Chunks and vectors are stored in IndexedDB under the file's hash. Drop the same PDF again and it loads instantly.
Finds what matters
Your question becomes another vector, compared by cosine similarity: the five closest chunks are the context.
Answers with citations
A small model answers using only those chunks, streaming token by token, and cites pages like this: [p. 12]. If it isn't in the document, it says so.

Models and requirements

Downloaded from Hugging Face the first time you drop a PDF (not when the page opens) and cached by the browser.

Embeddings≈ 120 MB
Xenova/multilingual-e5-small
384 dimensions, multilingual. 8-bit quantized.
Generation≈ 400 MB
onnx-community/Qwen2.5-0.5B-Instruct
Good Spanish for its size. 4-bit quantized.
Generation (fast)≈ 250 MB
HuggingFaceTB/SmolLM2-360M-Instruct
Lighter, weaker in Spanish. For low-memory machines.

Recommended: Chrome or Edge with WebGPU and 4 GB of free RAM. Without WebGPU it runs on WebAssembly, but generation is much slower.

FAQ

Does it really upload nothing?

Nothing. Open your browser's Network tab: during the first load you'll see downloads from huggingface.co (the model weights) and after that, no requests at all. The PDF is read in memory and never leaves your machine.

What gets downloaded and how big is it?

Two models: one for embeddings (about 120 MB) and one for generation (about 400 MB, or 250 with the fast option). They download once and stay in the browser cache.

Does it work without WebGPU?

Yes, on WebAssembly. Reading and search are just as fast; generating the answer is considerably slower. Firefox and Safari usually fall in this case.

Does it work with scanned PDFs?

Not yet. A scan is an image with no text to extract; it would need OCR. It's on the list for a second version.

How big a PDF can it handle?

Hundreds of pages, no problem: 300 pages are about 1,500 vectors and search is instant. What takes time is the first indexing, and that depends on your machine.

Where is everything stored and how do I delete it?

Models in the browser cache, indexed documents in IndexedDB and chats in localStorage; all on your machine. "Clear everything" wipes it all.

Does it make mistakes?

Sometimes. It's a 500-million-parameter model, not a cloud service: it can mix up facts that sit close together in the text. The [p. N] citations are there so you can check.

Why Ícaro?

Icarus flew too close to the sun. This one never goes up to the cloud.

Made byJulián·source·MIT

No analytics, no cookies, no server. The only downloads are the models, and only when you drop a PDF.

Ícaro

Private AI in your browser

Ícaro, your PDF assistant

Fly low: your PDF never goes to the cloud.

Drop a PDF, ask in your own words and get answers that cite the pages. Everything happens in your browser: neither the document nor your questions leave your machine.

No server
No account
No analytics
Open source