Agentic RAG Chatbot Demo

A demo chatbot that combines Retrieval-Augmented Generation (RAG) with agentic decision-making.
It answers questions about a local dataset, and can also query Yahoo Finance for real-time commodity and index prices.

The system is built with:

• LangGraph → to orchestrate agent flow
• FastAPI → RESTful backend
• FAISS → local vector store for document retrieval
• yfinance → live financial data
• HTML/CSS/JS frontend → simple chat UI (WhatsApp-style)

---

LangGraph Workflow

Flow Explanation:

1. Start → query_evaluate node
   • Decide whether the question is about a ticker price.
2. If finance query →
   • extract_ticker → yahoo_search → generate
3. If knowledge query →
   • retrieve → evaluate_documents → generate
4. End → returns answer

---

Architecture Overview Diagram

---

Features

• Retrieval from local JSON dataset of geopolitical/supply chain events
• Yahoo Finance price lookup for futures, indices, ETFs, large-cap stocks
• LangGraph workflow to decide:
  • “Is this a finance query?” → extract ticker + query Yahoo Finance
  • “Or a knowledge query?” → retrieve documents → generate answer
• Frontend chat window for interaction, with typing animation to signal inference

---

Project Structure

AGENTIC-RAG-DEMO/
├─ agent/
│  └─ app/
│     ├─ graph/                 # LangGraph wiring
│     │  ├─ build.py            # builds/compiles the workflow
│     │  └─ graph_chain.py      # shared graph/state helpers
│     ├─ nodes/                 # node functions (RAG + finance path)
│     │  ├─ evaluate_documents.py
│     │  ├─ evaluate_query.py
│     │  ├─ extract_state.py
│     │  ├─ generate.py
│     │  ├─ retrieve.py
│     │  └─ yahoo_finance_state.py
│     ├─ services/              # LLM services (structured outputs, routers, etc.)
│     ├─ tools/                 # utilities/clients
│     │  ├─ embed_texts.py      # OpenAI embedding wrapper
│     │  └─ yahoo_finance_api.py# yfinance wrapper
│     ├─ static/
│     │  └─ index.html          # chat UI (served by FastAPI)
│     ├─ main.py                # local runner for the graph
│     └─ server.py              # FastAPI app (/, /chat)
├─ data/
│  ├─ local.json                # source events
│  └─ vector_store/             # FAISS artifacts
│     ├─ ****.faiss
│     └─ ***_fused.json
├─ workflow_graph.png           # LangGraph diagram
├─ scripts/
│   ├─ build_knowledge_base.py  # build vector store
├─ .env.example
├─ .gitignore
├─ requirements.txt
└─ README.md

---

Running the Project

1. Install dependencies

pip install -r requirements.txt

2. Set environment variables

Create a .env file in the project root:

OPENAI_API_KEY=sk-xxxx...
QUERY_EXTRACTOR_MODEL=gpt-5
QUERY_EVAL_MODEL=gpt-5

3. Start the backend

uvicorn agent.app.server:app --reload

4. Open frontend

Visit http://localhost:8000 in your browser. Use the chat box to ask questions like:

• “What is the tariff situation between the US and the EU?”
• “What is the price of gold yesterday?”

Next Steps

• Stream answers token-by-token for smoother UX
• Deploy to cloud (e.g., GCP or AWS) with managed FAISS store
• Expand Yahoo Finance tool with more robust ticker recognition
• Expand the graph nodes/edges for more types of questions

License

MIT