Feature/langgraph gemini example

[ParthSingh0506]

Summary

Adds a complete LangGraph + Google Gemini integration example demonstrating APO (Automatic Prompt Optimization) training.

Features

• LangGraph integration with Google Gemini 2.5 Flash (non-OpenAI LLM)
• Bulletproof placeholder handling to prevent KeyErrors from APO's creative prompt generation
• Persistent training result storage (JSON, TXT, PKL formats)
• Windows compatibility fixes for path handling
• Comprehensive README with installation, troubleshooting, and customization guides

What's Included

• agent_train.py - Complete training script with news agent + APO setup
• README.md - Detailed setup instructions and troubleshooting
• requirements.txt - All Python dependencies
• .env.example - API key configuration template
• .gitignore - Prevents accidental secret commits

Problem Solved

This example addresses common issues developers face when integrating LangGraph with Agent Lightning:

1. KeyError on unexpected placeholders - APO's GPT-4 editor creates prompts with placeholders beyond {query}, causing runtime errors
2. Infinite training loops - Missing beam_rounds parameter causes training to run indefinitely
3. Lost optimization results - Incorrect prompt extraction methods lose the optimized prompt

Testing

Tested on:

• ✅ Windows 11 with Python 3.12
• ✅ Standard pip + venv workflow
• ✅ Successfully completes 3 training rounds
• ✅ Produces optimized prompt with improved validation scores

Related

This example fills a gap in the current examples - there are no LangGraph integrations or non-OpenAI LLM examples in the repository yet.

Note: I'm writing a companion blog article with a detailed walkthrough that
can serve as additional community documentation.

[ParthSingh0506]

@microsoft-github-policy-service agree