📘 PDF Chat Application (RAG with Highlighted Sources)

This project is a Streamlit-based PDF Question Answering app built using Retrieval-Augmented Generation (RAG).
You can upload a PDF, ask questions about its content, and get short, accurate answers generated by Gemini, along with highlighted source chunks from the PDF.

---

🚀 Features

• 📄 Upload any PDF file
• 🔍 Semantic search using FAISS
• 🧠 Text embeddings via HuggingFace (MiniLM)
• 🤖 Answer generation using Google Gemini
• 🖍️ Highlighted source chunks with page numbers
• ⚡ Cached vector store for fast performance
• ❌ Responds with "I don't know" if answer is not in the PDF
• 🧾 Stores Chat History

---

🛠️ Tech Stack

• Streamlit – UI
• LangChain – Orchestration & document processing
• FAISS – Vector database
• HuggingFace Embeddings – Semantic text embeddings
• Google Gemini – Answer generation
• PyPDFLoader – PDF extraction

---

📂 Project Structure

├── main3.py # Streamlit app
├── .env # API keys
├── requirements.txt # Dependencies
└── README.md

---

📦 Installation

pip install -r requirements.txt

🧠 How It Works

• Upload a PDF – Use the file uploader widget
• PDF Processing – Document is split into overlapping chunks (800 chars, 150 char overlap)
• Embeddings – Chunks are converted into embeddings using HuggingFace MiniLM
• Vector Store – Embeddings are stored in FAISS for fast retrieval
• Query Processing – User question retrieves top-3 relevant chunks
• Answer Generation – Gemini generates an answer strictly from retrieved context
• Source Display – Original chunks are shown with page numbers and highlighting

---

🧪 Example Prompts

Try asking these questions after uploading a PDF:

• "What is the main topic discussed in this PDF?"
• "Summarize the key points"
• "What does the document say about [specific topic]?"
• "Who are the authors?"

---

⚠️ Important Notes

Accuracy & Grounding

• ✅ Answers are strictly grounded in PDF context
• ✅ No hallucinations – If information is missing, the app responds with "I don't know"
• ❌ Does not generate information outside the PDF content

PDF Compatibility

• ✅ Works best with text-based PDFs
• ❌ May struggle with scanned images (OCR not included)

Performance

• Vector stores are cached for faster repeated queries
• First query may take a few seconds while embeddings are computed