FlightGPT — Conversational Flight Search Bot

A conversational Q&A system that lets you search flight information and trajectories using natural language. It combines a Large Language Model with a Neo4j graph database to translate plain English questions into Cypher queries, returning answers through text, interactive maps, data tables, and KPI statistics.

---

Features

• Natural Language Queries — Ask questions like "Show me all flights from Cologne to Paris" without writing any database queries
• Interactive Map — Flight trajectories and airports visualized on an ArcGIS satellite map
• Streaming Responses — Real-time status updates while queries are processing
• Persona System — Choose from 5 response styles, from casual to military-precise
• GeoJSON Spatial Filter — Upload a polygon to find flights passing through a custom area
• Multi-Modal Output — Answers come with a text response, map data, HTML tables, and KPI cards

---

Tech Stack

Layer	Technology
Frontend	HTML5, ArcGIS Maps SDK for JavaScript, Calcite Components
Backend	Flask
LLM Orchestration	LangGraph + LangChain
LLM Provider	GWDG OpenAI-compatible API
Database	Neo4j (Aura Cloud)
Geospatial	Shapely

---

Getting Started

Prerequisites

• Python 3.10+
• Access to a Neo4j Aura instance with flight data loaded
• A GWDG API key (or compatible OpenAI endpoint)

Installation

git clone https://github.com/MatteoWeickert/flight-search-bot.git
cd flight-search-bot
pip install -r requirements.txt

Environment Configuration

Create a .env file in the project root:

NEO4J_URI=neo4j+s://<instance>.databases.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=<password>
NEO4J_DATABASE=neo4j

GWDG_API_KEY=<api_key>
GWDG_BASE_URL=https://chat-ai.academiccloud.de/v1
GWDG_MODEL=openai-gpt-oss-120b

Run

cd src
python main.py

Then use the LiveServer Extension (index.html) within VSC for viewing the frontend.

---

How It Works

1. Query Refinement — Follow-up questions are rewritten into standalone queries using conversation history
2. Routing — An LLM router decides which agents to invoke (text answer, map, table)
3. Cypher Generation — The QA agent translates the question into a Cypher query
4. Execution — The query runs against Neo4j and results are returned
5. Output — Results are formatted as a text answer, GeoJSON for the map, an HTML table, and KPI cards

The orchestration is built with LangGraph, giving the workflow a deterministic, auditable structure.

---

Project Structure

flight-search-bot/
├── src/
│   ├── main.py                 # Flask server
│   ├── agents/
│   │   ├── supervisor.py       # LangGraph workflow
│   │   ├── qa_agent.py         # Text answers + Cypher generation
│   │   ├── map_agent.py        # GeoJSON map output
│   │   └── table_agent.py      # HTML tables + KPI cards
│   ├── utils/
│   │   └── geojson_input.py    # Spatial filtering with Shapely
│   └── application/
│       ├── index.html          # UI layout
│       ├── main.js             # Frontend logic
│       └── style.css           # Styling
├── data/                       # Sample CSV data (100 flights)
├── requirements.txt

---

Data Model

Flight data is stored in Neo4j as a graph. Trajectories are modelled as a linked list of TrajectoryPoint nodes connected by NEXT relationships, with the Flight node pointing to the first point via HAS_POINT.

Key node types: Flight, Airport, TrajectoryPoint, FIR (Flight Information Region)

---

Limitations

• Dataset is currently limited to 100 sample flights
• No user authentication
• LLM-generated Cypher queries may occasionally fail on edge cases
• Single LLM provider (GWDG/OpenAI-compatible API)