Streaming Subgraph Demo (LangChain + LangGraph)

This repo contains a minimal example that demonstrates how to preserve token-by-token streaming when calling a LangGraph subgraph. The script contrasts a "broken" pattern vs the exact "updated" pattern that uses stream_mode="messages".

What's inside

• new_streaming_example.py: runnable demo comparing broken vs fixed streaming behavior
• requirements.txt: Python dependencies (LangChain, LangGraph, OpenAI SDK, dotenv)

Prerequisites

• Python 3.10+ recommended
• An OpenAI API key with access to the model used in the script (defaults to gpt-4o-mini)

Setup

1. Create and activate a virtual environment

• macOS/Linux:

python3 -m venv .venv
source .venv/bin/activate

• Windows (PowerShell):

py -m venv .venv
.venv\Scripts\Activate.ps1

2. Install dependencies

pip install -r requirements.txt

3. Provide environment variables

The script loads environment variables from a local .env file if present. Create a file named .env in the repository root with your keys:

# Required
OPENAI_API_KEY=your-openai-api-key

# Optional LangSmith (for tracing/observability)
LANGSMITH_PROJECT=default
LANGSMITH_API_KEY=your-langsmith-api-key
LANGCHAIN_TRACING_V2=true

Never commit real secrets. If you're using git, ensure .env is ignored.

Run the demo

python new_streaming_example.py

You should see two sections:

• BROKEN: shows the non-streaming behavior when invoking the subgraph directly
• FIXED: shows chunked streaming when using stream_mode="messages"

Troubleshooting

• "Module not found": Ensure your virtual environment is activated and dependencies are installed.
• No streaming observed in the FIXED path: Confirm your langgraph version is >= 0.6.4.

  python -c "import langgraph; print(langgraph.__version__)"

• Authentication errors: Verify OPENAI_API_KEY is set in .env or your shell environment and that your key has access to the chosen model.

Notes

• This example mirrors the fix described by LangGraph maintainers for preserving streaming from subgraphs by explicitly requesting messages streaming: subgraph.stream(state, stream_mode="messages").