🛡️ Insurance FAQ Chatbot

A local Retrieval-Augmented Generation (RAG) chatbot that answers insurance questions based on your own PDF documents. Built with TinyLlama, FAISS, and Gradio.

---

User Experience

PDFs are static: users have to read and search manually. Chatbot is interactive: users can ask natural language questions and get direct answers.

For insurance FAQs, this means: “What’s the claim process for car insurance?” → instant answer. No scrolling, no guessing, no searching multiple pages.

How It Works

1. PDF Ingestion — loads and chunks your insurance FAQ document
2. Embedding — converts chunks into vector embeddings using all-MiniLM-L6-v2
3. Vector Store — stores embeddings in a FAISS index for fast similarity search
4. RAG Retrieval — on each query, retrieves the top-3 most relevant chunks
5. LLM Generation — feeds retrieved context + question to TinyLlama to generate an answer
6. Gradio UI — serves a chat interface accessible via browser

User Query
    │
    ▼
Embedding Model ──► FAISS Index ──► Top-K Chunks
                                         │
                                         ▼
                                   TinyLlama LLM
                                         │
                                         ▼
                                      Answer

---

Project Structure

llm_insurance-chatbot/
├── main.py                  # Main application
├── requirements.txt         # Python dependencies
├── data/
│   └── Insurance_FAQs.pdf   # Your source document (add this yourself)
└── embeddings/
    └── vector_store.pkl     # Auto-generated on first run

---

Setup

1. Clone / download the project

cd llm_insurance-chatbot

2. Create a virtual environment (recommended)

python -m venv venv

# Windows
venv\Scripts\activate

# macOS/Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Add your PDF

Place your insurance FAQ document at:

data/Insurance_FAQs.pdf

5. Run the app

python main.py

On first run, the vector store will be built automatically. Subsequent runs will load it from cache.

---

Accessing the UI

After running, you'll see output like:

* Running on local URL:  http://127.0.0.1:7860
* Running on public URL: https://xxxxxx.gradio.live

• Local URL — open in your browser on the same machine
• Public URL — share with anyone for 72 hours (enabled via share=True)

---

System Requirements

Component	Minimum
RAM	6GB free
Python	3.10+
Disk	~3GB (for model cache)
GPU	Optional (runs on CPU)

Note for Windows users: You may see a symlinks warning from HuggingFace. This is harmless. To fix it, enable Developer Mode in Windows Settings or run Python as Administrator.

---

Configuration

Key parameters you can tweak in main.py:

Parameter	Location	Default	Description
chunk_size	chunk_text()	500	Characters per chunk
overlap	chunk_text()	50	Overlap between chunks
top_k	retrieve_relevant()	3	Number of chunks retrieved
max_new_tokens	generate_answer()	300	Max response length
temperature	generate_answer()	0.7	Response creativity (0=deterministic)

---

Troubleshooting

Segmentation fault during model load → Not enough RAM. The model needs ~4GB free. Close other applications and retry.

localhost is not accessible error → Add share=True to iface.launch() and use the public URL instead.

Slow responses → Expected on CPU — TinyLlama takes 30–90 seconds per response without a GPU.

Poor answer quality → Check that your PDF text is extractable (not a scanned image). Try increasing top_k to 5.

---

Dependencies

Package	Purpose
pdfplumber	PDF text extraction
sentence-transformers	Generating embeddings
faiss-cpu	Vector similarity search
transformers	Loading TinyLlama LLM
torch	Model inference backend
gradio	Chat web interface
accelerate	Optimized model loading