✈️ Flight Agent

A conversational AI agent for flight booking and KLM baggage policy queries. Built with LangGraph's ReAct loop — the agent decides which tool to call based on what you ask, remembers your conversation across messages, and responds in plain English.

Live demo → luxury-youtiao-5bef59.netlify.app
API → flight-agent-1.onrender.com

⚠️ Demo project. Uses hardcoded flight data, not real airline APIs. The RAG tool retrieves from a local KLM baggage policy document — not a live search tool.

---

What it does

• Check flight availability between airports
• Book a flight (returns a mock confirmation)
• File a customer complaint
• Answer KLM baggage policy questions — powered by BM25 retrieval over klm_baggage.txt

---

Tech Stack

Layer	Technology
Agent framework	LangGraph (ReAct loop)
LLM	Groq — llama-3.3-70b-versatile
RAG retrieval	BM25Retriever (rank_bm25) — no local ML model
Conversation memory	LangGraph MemorySaver checkpointer
Backend	FastAPI + Uvicorn, deployed on Render
Frontend	Vanilla HTML/CSS/JS, deployed on Netlify
Cross-origin	FastAPI CORSMiddleware

Why BM25 instead of vector embeddings?
The original implementation used HuggingFace sentence-transformers which pulls in PyTorch — blowing past Render's free tier 512MB RAM limit on startup. BM25 is pure Python, needs ~5MB RAM, and works well for keyword-heavy policy queries.

---

Architecture

User → Netlify frontend
         ↓ POST /chat
    FastAPI (Render)
         ↓
    LangGraph ReAct agent
         ↓ (tool routing)
    ┌────────────────────────────┐
    │  check_flight              │
    │  book_flight               │
    │  book_complaint            │
    │  rag_tool (BM25 retriever) │
    └────────────────────────────┘
         ↓
    Groq API (llama-3.3-70b)

The agent uses a MemorySaver checkpointer — each session gets a thread_id so the agent remembers earlier messages within the same conversation.

See ARCHITECTURE.md for a full breakdown of each file.

---

File Structure

├── api.py           # FastAPI app — /chat endpoint + CORS
├── graph.py         # LangGraph StateGraph + MemorySaver checkpointer
├── tools.py         # BM25 RAG tool + flight/complaint tools
├── context.py       # Model name + system prompt config
├── state.py         # State and InputState definitions
├── utils.py         # load_chat_model helper
├── prompts.py       # System prompt
├── klm_baggage.txt  # RAG knowledge base (KLM baggage policy)
├── requirements.txt
└── runtime.txt      # Python version pin for Render

---

Run locally

git clone https://github.com/MAGUIRE-GOATED/Flight_agent
cd Flight_agent
python -m venv venv && source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

Create a .env file:

GROQ_API_KEY=your_key_here

Start the API:

uvicorn api:app --reload

Then open http://localhost:8000/docs to try it via Swagger UI.

---

Deployment

• Backend — Render (free tier). runtime.txt pins the Python version. Auto-deploys on push to main. Note: free instances sleep after inactivity — first request after idle takes ~50s.
• Frontend — Netlify (drag and drop index.html). CORS is scoped to the Netlify URL in api.py.