🧠 Multi-Query + Reranking RAG PDF Chatbot

This project is a Streamlit-based RAG (Retrieval-Augmented Generation) PDF chatbot that uses:

• UV for dependency management
• PyMuPDF for PDF extraction
• Sentence Transformers for embeddings
• FAISS for vector search
• Multi-query expansion using DeepSeek
• Cross-Encoder reranking
• Ollama for LLM-based answering

---

🚀 Features

✔ Extracts text from PDF
✔ Splits text into chunks
✔ Creates a FAISS vector index
✔ Performs multi-query retrieval
✔ Re-ranks retrieved chunks
✔ Generates answers using context
✔ Uses DeepSeek-R1 (via Ollama)

---

📦 Project Setup

Follow these steps to run the application:

---

1️⃣ Clone the Repository

git clone <repo-link>
cd rag_tutorials

---

2️⃣ Install Dependencies Using UV

Make sure UV is installed:

pip install uv

Then run:

uv sync

This will install all dependencies listed in pyproject.toml.

---

3️⃣ Run the Streamlit Application

uv run streamlit run main.py

---

📁 File Structure

project-folder/
│── main.py
│── pyproject.toml
│── README.md

---

⚙️ Requirements

Your pyproject.toml should include dependencies such as:

streamlit
pymupdf
faiss-cpu
sentence-transformers
numpy
ollama
cross-encoder

UV will automatically install them.

---

📝 Description of the Workflow

1. PDF Extraction

Uses PyMuPDF to extract text from each page.

2. Chunking

Splits PDF text into overlapping chunks.

3. Vector Embedding

Embeds chunks using all-MiniLM-L6-v2.

4. FAISS Index

Stores embeddings in an IndexFlatL2 vector store.

5. Multi-Query Retrieval

DeepSeek R1 reformulates user queries into variations.

6. Re-ranking

Uses CrossEncoder (MS MARCO MiniLM) to score chunks.

7. LLM Response

Ollama generates final answer based on top-ranked context.

---