Weather Agent (LangGraph + Bedrock Mantle)

A LangGraph ReAct agent backed by Amazon Bedrock Mantle via its OpenAI-compatible API. It answers weather questions using a get_weather tool (powered by wttr.in).

LangChain is used only for primitives (ChatOpenAI, @tool). The agent loop is a compiled LangGraph from langgraph.prebuilt.create_react_agent.

Prerequisites

• Python 3.10+ (local dev) or Docker
• A Bedrock Mantle API key and base URL for your region

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Create a .env file in the project root:

OPENAI_API_KEY=bedrock-api-key-***
OPENAI_BASE_URL=https://bedrock-mantle.ap-south-1.api.aws/v1
OPENAI_MODEL=openai.gpt-oss-20b
PORT=80

Default model is openai.gpt-oss-20b — the cheapest Bedrock Mantle option (~$0.07/1M input, ~$0.30/1M output vs 2× for 120b).

Run locally (CLI)

python agent.py

Run locally (HTTP API)

uvicorn app:app --reload --port 8000

curl http://localhost:8000/health

curl -X POST http://localhost:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "What is the weather in Paris?"}'

Interactive API docs: http://localhost:8000/docs

Run with Docker

Create a .env file with your Bedrock credentials, then:

docker compose up --build

The service listens on http://localhost:80 (host and container both use port 80).

curl -X POST http://localhost/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Is it warm in Mumbai right now?"}'

Stop the service:

docker compose down

Project layout

agent.py            # LangGraph agent factory, invoke logic, CLI
app.py              # FastAPI HTTP server
session_store.py    # In-memory sessions (not LangGraph checkpointing)
tools.py            # Weather tool
Dockerfile
docker-compose.yml
requirements.txt

How it works

1. ChatOpenAI calls Bedrock Mantle at /v1/chat/completions.
2. create_react_agent (LangGraph) runs a ReAct loop: the model may call tools, then answer.
3. get_weather fetches live data from wttr.in and returns a short summary.
4. Sessions use an in-app SessionStore keyed by thread_id. Each HTTP/CLI turn invokes the graph with only the current user message; recent cities from prior turns are injected as short context. This is not LangGraph’s checkpointer — the graph does not persist message history across invokes.

Sessions

Each conversation is keyed by a thread_id. Reuse the same ID so follow-up questions can use session context (e.g. recently mentioned cities).

• First message without thread_id → server creates one and returns it
• Follow-up messages → send the same thread_id back
• Sessions live in memory inside the process (lost on restart)

Example multi-turn flow:

# Start a session
curl -X POST http://localhost/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "What is the weather in Paris?"}'
# → {"reply": "...", "thread_id": "abc-123"}

# Continue the same session
curl -X POST http://localhost/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "What about London?", "thread_id": "abc-123"}'

# End a session (optional)
curl -X DELETE http://localhost/sessions/abc-123

With docker-compose, each container has its own SessionStore. For multiple replicas, use a shared store (e.g. Redis) or adopt a LangGraph checkpointer backed by Redis/Postgres if you want full graph state persistence.

Session-based guardrails (Akto gateway)

When the caller sends an x-session-id header on /chat, it is forwarded on the LLM calls so the Akto gateway can apply session-based guardrails. If the caller omits the header, no session header is sent upstream:

curl -X POST http://localhost/chat \
  -H "Content-Type: application/json" \
  -H "x-session-id: my-session-1" \
  -d '{"message": "What is the weather in Paris?"}'

API

Method	Path	Description
GET	/health	Liveness check
POST	/chat	Send a message, get a reply
DELETE	/sessions/{thread_id}	Clear a conversation session

POST /chat body:

{
  "message": "What's the weather in Paris?",
  "thread_id": "optional-existing-session-id"
}

Response:

{
  "reply": "...",
  "thread_id": "abc-123"
}