# 📄 PDF Chatbot — Runs in Your Browser

A **zero-server, zero-API-key** chatbot that lets you ask questions about your PDFs using lightweight AI models running entirely in your browser.

🔗 **Live Demo**: https://huggingface.co/spaces/ghostofpk20/pdf-chatbot-browser

---

## 🚀 Try It Now (No Installation!)

Just open the link above and start chatting with your PDFs. That's it.

> **First use only**: The AI embedding model (~22MB) downloads from Hugging Face. After that, it works **completely offline**.

---

## 🧠 How It Works

| Step | What Happens |
|------|-------------|
| 1. Upload PDF | PDF.js extracts text entirely in your browser |
| 2. Chunk & Index | Text is split into chunks; each chunk gets an AI embedding vector |
| 3. Ask Question | Your question is embedded; top-K matching chunks are retrieved |
| 4. Get Answer | Best sentences from matching chunks are extracted and returned |

All processing happens **locally** — your PDFs never leave your device.

---

## 🎯 Key Features

- ✅ **Fully offline** after first download — no servers, no API keys
- ✅ **Drag & drop** PDF upload
- ✅ **Multiple PDFs** — chat with any uploaded document
- ✅ **WebGPU acceleration** — 5-10x faster on Chrome/Edge with GPU
- ✅ **WASM fallback** — works on any modern browser
- ✅ **Keyword fallback** — still searches if AI model fails
- ✅ **Sources shown** — every answer includes the document excerpts used
- ✅ **No scanned PDFs** — text-based PDFs only (browser limitation)

---

## 📦 Chrome Extension (Bonus)

Download `chrome-extension.zip` from the files tab and load it as an unpacked extension in `chrome://extensions/`.

---

## 🛠️ Models Used

| Model | Size | Purpose |
|-------|------|---------|
| `Xenova/all-MiniLM-L6-v2` | ~22MB (q8) | Text embeddings for document retrieval |

- **ONNX quantized** to 8-bit for minimal memory
- **Runs via Transformers.js** + ONNX Runtime Web
- **Answers are extractive** — no LLM hallucinations, just facts from your document

---

## 🔒 Privacy

- **Zero network calls** after model download
- **Your PDFs never leave your browser**
- **No analytics, no tracking, no API keys**
- Everything stays in browser memory only

---

## 📋 Browser Support

| Browser | Status |
|---------|--------|
| Chrome 113+ | ✅ Full (WebGPU) |
| Edge 113+ | ✅ Full (WebGPU) |
| Firefox | ⚠️ WASM fallback |
| Safari 17.4+ | ⚠️ WebGPU in preview |

---

## 🙏 Credits

- [Transformers.js](https://huggingface.co/docs/transformers.js) — ML in the browser
- [PDF.js](https://mozilla.github.io/pdf.js/) — PDF text extraction
- [ONNX Runtime Web](https://onnxruntime.ai/) — Model inference