A Retrieval-Augmented Generation (RAG) chatbot built to answer customer support queries using Angel One’s documentation and insurance PDFs.
Powered by FAISS for vector search and a lightweight LLM (TinyLlama) for local inference.
.jpg)
FEATURES
- Answers user queries using real support documents
- Uses FAISS vector store for efficient retrieval
- Semantic chunk search using SentenceTransformers
- Local LLM (TinyLlama or Mistral 7B via GGUF or Hugging Face Transformers)
- Gradio UI for an interactive chatbot interface
- Handles both general support topics and insurance-related queries
LIVE DEMO
Try it live: https://huggingface.co/spaces/harshitsingh911/alltius
TECH STACK
Component - Technology
UI - Gradio
Model - TinyLlama (transformers pipeline)
Embeddings - SentenceTransformers (all-MiniLM-L6-v2)
Vector Store - FAISS
Backend - Python
HOW IT WORKS
- Documents Loading: PDFs (Angel One support docs + insurance info) are loaded and chunked into manageable segments.
- Embedding & Indexing: Chunks are converted into dense embeddings and stored using FAISS.
- Retrieval: On user query, top relevant chunks are fetched from FAISS using cosine similarity.
- Prompt Construction: The selected chunks are added to the query and sent as a prompt to the LLM.
- Response Generation: The LLM generates an answer based only on the retrieved documents.
SAMPLE QUERIES
- What is SIP?
- How do I reset my trading password?
- How to buy insurance using Angel One?
- What are brokerage charges?
RUN LOCALLY
- Clone the repo
git clone https://github.com/harshitsingh09/RAG-Chatbot.git cd angelone-rag-bot
- Install dependencies
pip install -r requirements.txt
- Run the app
python app.py
This will launch the Gradio app at http://127.0.0.1:7860
FILE STRUCTURE
.
├── app.py -> Gradio frontend + chatbot logic
├── requirements.txt -> Python dependencies
├── vectorstore.index -> FAISS index file
├── metadata.pkl -> Chunk metadata for retrieval
├── README.md -> Project info and guide