📚 PDF RAG Chatbot

A fully local Retrieval-Augmented Generation (RAG) chatbot that lets you upload PDF documents and ask questions about their contents — no external APIs, no cloud dependencies, no API keys required.

---

How It Works

Upload PDFs → Extract Text → Chunk → Embed → FAISS Index → Question → Retrieve → LLM → Answer

1. PDFs are parsed page-by-page and split into overlapping word chunks
2. Each chunk is converted into a semantic vector using all-MiniLM-L6-v2
3. Normalized embeddings are stored in a local FAISS index (cosine similarity via inner product)
4. When you ask a question, the top-5 most relevant chunks are retrieved
5. FLAN-T5 generates a document-grounded answer using those chunks + chat history as context

---

Features

• 📄 Upload and process multiple PDFs in one go
• 🔍 Semantic similarity search via FAISS (IndexFlatIP)
• 🧠 Fully local LLM inference with FLAN-T5 — no API keys required
• 💬 Persistent chat history with multi-turn context
• 📚 "View Retrieved Context" expander showing source file, page number, and similarity score
• 💾 Index persists to disk — no need to re-upload documents on restart
• 🗑️ One-click index and document clearing
• ⚡ Model caching with @st.cache_resource for fast reloads

---

Tech Stack

Component	Technology
UI	Streamlit
PDF Parsing	PyPDF2
Text Embeddings	sentence-transformers/all-MiniLM-L6-v2
Vector Store	FAISS (faiss-cpu)
Language Model	google/flan-t5-base (HuggingFace Transformers)
Numerical Processing	NumPy

---

Project Structure

├── app.py                   # All logic in one file — PDF processing, chunking,
│                            # embedding, FAISS index, LLM, and Streamlit UI
├── requirements.txt
├── faiss_store/             # Auto-created after first index build
│   ├── index.faiss          # Persisted FAISS vector index
│   └── meta.pkl             # Persisted chunk metadata
└── uploaded_documents/      # Auto-created when PDFs are uploaded

---

Getting Started

Prerequisites

• Python 3.8+
• ~1–2 GB disk space for model downloads on first run

Installation

1. Create and activate a virtual environment

# Windows
python -m venv venv
venv\Scripts\activate

# macOS / Linux
python -m venv venv
source venv/bin/activate

2. Install dependencies

pip install -r requirements.txt

3. Run the app

streamlit run app.py

Opens automatically at http://localhost:8501.

---

Usage

Step 1 — Upload PDFs Use the sidebar file uploader to add one or more PDF documents.

Step 2 — Build Index Click 🔨 Build Index. The app will:

• Save uploaded files to uploaded_documents/
• Extract text page-by-page
• Split into overlapping word chunks
• Generate and normalize embeddings
• Store everything in a FAISS index on disk

Step 3 — Ask Questions Type any question in the chat input at the bottom. The system retrieves the most relevant passages and generates a grounded answer.

Step 4 — Inspect Sources Expand 🔎 View Retrieved Context below any answer to see which file and page each chunk came from, along with its similarity score.

Clearing the Index Click 🧹 Clear Index in the sidebar to wipe the FAISS index, metadata, and all uploaded documents.

---

Configuration

All parameters are defined at the top of app.py and can be changed there:

Parameter	Variable	Default
Embedding model	EMBED_MODEL	sentence-transformers/all-MiniLM-L6-v2
Language model	LLM_MODEL	google/flan-t5-base
Chunk size	CHUNK_WORDS	500 words
Chunk overlap	CHUNK_OVERLAP	80 words
Top-K retrieval	TOP_K_DEFAULT	5 chunks
Max context length	MAX_CONTEXT_CHARS	6000 characters
Max generated tokens	MAX_NEW_TOKENS	600

---

How the Index Works

Embeddings are L2-normalized before being added to a faiss.IndexFlatIP index. Because the vectors are unit-normalized, inner product search is mathematically equivalent to cosine similarity — giving semantically meaningful rankings without a dedicated cosine index.

The index and its metadata are saved to faiss_store/ after every build, so the app reloads them automatically on restart without reprocessing your documents.

---

Notes

• All processing and inference runs locally — your documents never leave your machine.
• FLAN-T5 is best for factual, extractive answers. For richer responses, swap LLM_MODEL for a larger model such as google/flan-t5-large.
• If a question cannot be answered from the retrieved context, the model is explicitly instructed to say so rather than hallucinate.
• Scanned PDFs (image-only) will not yield extractable text — use OCR-preprocessed PDFs for those.