Enterprise Agentic AI Platform

A compact reference implementation for an enterprise-style agentic AI workflow built with FastAPI, LangGraph, Qdrant, Neo4j, and an OpenAI-compatible reasoning layer.

What It Does

The platform routes an incoming operational question through specialized agents:

• router_agent selects the branch to run
• retrieval_agent queries vectorized operational documents in Qdrant
• tool_agent calls the weather MCP service
• graph_agent queries Neo4j for graph context
• reasoning_agent produces the final recommendation

Current routing behavior:

• Weather and flight-impact queries go through the weather branch
• General policy and operations queries go through the RAG branch

Stack

Layer	Technology
API	FastAPI
Workflow orchestration	LangGraph
Reasoning	OpenAI-compatible chat completions
Vector store	Qdrant
Graph store	Neo4j
Embeddings	SentenceTransformers
Tool service	FastAPI weather MCP
Runtime	Docker Compose

Repository Layout

agentic-ai-platform-enterprise/
├── agents/
├── api/
├── rag/
├── tools/
├── workflows/
├── .env
├── .gitignore
├── docker-compose.yml
├── Dockerfile
├── README.md
└── requirements.txt

Configuration

The repo now includes a root .env file. The main settings are:

API_PORT=8000
OPENAI_API_KEY=
OPENAI_MODEL=gpt-4o-mini
EMBEDDING_MODEL=all-MiniLM-L6-v2
QDRANT_COLLECTION=airline_docs
QDRANT_URL=http://qdrant:6333
NEO4J_URI=bolt://neo4j:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=password
WEATHER_MCP_URL=http://weather-mcp:9001
DEFAULT_AIRPORT=ATL

Notes:

• OPENAI_API_KEY is optional for local testing
• If no OpenAI key is set, the app returns a deterministic fallback recommendation
• API_PORT controls the host port exposed by Docker Compose

Quick Start

1. Start the stack:

docker compose up --build -d

2. Ingest the sample RAG documents into Qdrant:

docker compose exec api python rag/ingest_docs.py

3. Query the API:

http://localhost:8000/query?q=Storm expected in ATL at 17:00

If port 8000 is already in use, set API_PORT in .env before starting Compose.

API Response Shape

GET /query?q=... returns a structured response:

{
  "query": "Storm expected in ATL at 17:00",
  "route": "weather",
  "decision": "Monitor ATL operations due to high storm conditions.",
  "weather": {
    "airport": "ATL",
    "event": "Storm",
    "severity": "High"
  },
  "rag_context": null,
  "graph_data": [],
  "errors": []
}

For RAG-oriented queries, route will be rag and rag_context will contain retrieved document snippets.

Local Behavior Notes

• The weather service is internal to Compose and is called through WEATHER_MCP_URL
• Qdrant and Neo4j are networked internally in Compose and do not need host port bindings for normal app usage
• Neo4j graph results are currently best-effort; if the sample database has no Flight nodes, the app still returns a response
• The reasoning layer records upstream failures in the errors field instead of silently masking them

Example Queries

Weather route:

/query?q=Storm expected in ATL at 17:00

RAG route:

/query?q=What does the ops manual say about hub congestion?

Development Notes

• The Dockerfile installs dependencies before copying the rest of the source to improve build caching
• .gitignore excludes Python bytecode, virtual environments, local env files, and common cache directories
• .dockerignore keeps the build context smaller by excluding cache and local development artifacts

Author

Satish Gopinathan
Website: https://www.eagleeyethinker.com

License

MIT