Simple Workflow With LangGraph

This project is a minimal weather app built to show where LangGraph differs from plain LangChain.

Instead of letting a chain or agent decide everything in one implicit flow, this app makes the workflow explicit:

1. Classify the user's intent with an LLM.
2. Route execution to a graph node based on that intent.
3. Call a real weather tool backed by Open-Meteo only for weather requests.
4. Return a final response from the selected branch.

Why This Is LangGraph

If you already know LangChain, the key difference in this example is control flow.

• LangChain gives you the building blocks: prompts, models, tools, structured output.
• LangGraph adds explicit workflow orchestration: state, nodes, edges, and routing.

In this app:

• The intent classifier is a normal LangChain-style LLM call.
• The Open-Meteo weather tool is a normal LangChain tool.
• The router and branch execution are the LangGraph part.

That means you can see and control exactly how the program moves from one step to the next.

Reference Project

LangChain mirror project:

• simple-weather-app-with-langchain

Essential difference:

• This LangGraph project models control flow explicitly with state + nodes + conditional edges.
• The LangChain mirror project relies on an agent loop where tool-use control flow is implicit.

Workflow

START
  |
  v
classify_intent
  |
  +--> weather --> weather_node --> END
  |
  +--> chat ----> chat_node -----> END
  |
  +--> unknown -> unknown_node --> END

Requirements

• Python 3.9+
• MOONSHOT_API_KEY set in your environment

This demo uses:

• model: kimi-k2.5
• temperature: 1.0
• base URL: https://api.moonshot.cn/v1

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

This project is intended to run inside its local .venv only.

Run

python3 app.py

Example prompts:

• What's the weather in Shanghai tomorrow?
• Will it rain in Beijing?
• Hello
• Help me write a SQL query

File Layout

• app.py: thin entrypoint that starts the CLI
• config.py: runtime constants and environment lookup
• models.py: shared state and structured output schemas
• logging_utils.py: logging configuration and decorators
• llm.py: Moonshot client setup and LLM exchange logging
• weather.py: Open-Meteo integration and weather formatting
• graph.py: LangGraph nodes, routing, and graph assembly
• cli.py: interactive CLI loop
• doc/design.md: design notes and function-by-function explanation
• pyproject.toml: project metadata and dependencies

Notes

• The weather tool resolves the location with Open-Meteo Geocoding and then fetches current plus short-range forecast data from Open-Meteo Forecast.
• The LLM is only used for intent classification and location extraction.
• The routing logic is explicit and deterministic after classification.
• The app logs function inputs, return values, LLM request/response payloads, and Open-Meteo HTTP request/response payloads.
• Logging is off by default. Set APP_LOGGING=true to enable detailed traces.
• You can override the default endpoint with MOONSHOT_BASE_URL.
• kimi-k2.5 only accepts temperature 1, so this project defaults to 1.0.