Quick Start: Testing LangGraph with EvalView

Quick Start: Testing LangGraph with EvalView

This guide gets you testing your LangGraph agent in 5 minutes.

Step 1: Start Your LangGraph Agent

# Navigate to your LangGraph project
cd /path/to/your/langgraph-agent

# Start the server (check your project's README for the exact command)
python main.py
# OR
uvicorn main:app --reload --port 8000

# Verify it's running:
curl http://localhost:8000

Step 2: Connect EvalView to Your Agent

# Auto-detect and configure the endpoint
evalview connect

# Or manually specify your endpoint:
evalview connect --endpoint http://localhost:8000/api/chat

This will:

• ✅ Test the connection
• ✅ Auto-update .evalview/config.yaml
• ✅ Tell you the correct adapter type

Step 3: Run Your Tests

# Run with verbose output to see what's happening
evalview run --verbose

Step 4: Review Results

After running tests, you'll see:

• ✅ Pass/fail status for each test
• 📊 Scores, costs, and latency
• 💾 Results saved to .evalview/results/

View detailed report:

evalview report .evalview/results/LATEST.json --detailed

Troubleshooting

"Cannot connect to agent"

# Check if your agent is running
curl -X POST http://localhost:8000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"query": "test"}'

# If not working, check:
# 1. Is the server running?
# 2. Is it on port 8000?
# 3. What's the correct endpoint path?

"Wrong endpoint"

Update .evalview/config.yaml:

endpoint: http://localhost:YOUR_PORT/YOUR_PATH

"Tool names don't match"

Run with --verbose to see the actual API response, then update your test YAML files with the correct tool names.

Next Steps

1. Customize test cases in tests/test-cases/
2. Add more scenarios - Copy and modify the example YAMLs
3. Adjust thresholds - Based on your agent's actual performance
4. Run regularly - Add to your CI/CD pipeline

Example Test Case

Create tests/test-cases/my-test.yaml:

name: "My Custom Test"
description: "Test my agent does X"

input:
  query: "Your test query here"
  context: {}

expected:
  tools: []  # Add tool names after seeing verbose output
  output:
    contains:
      - "expected keyword 1"
      - "expected keyword 2"
    not_contains:
      - "error"

thresholds:
  min_score: 70
  max_cost: 0.50
  max_latency: 10000

Tips

• 🔍 Always use --verbose initially to understand API responses
• 📝 Start with loose thresholds, tighten based on actual performance
• 🧪 Test one scenario at a time when debugging
• 💰 Set OPENAI_API_KEY in .env.local for LLM-as-judge evaluation

Need help? Check DEBUGGING.md or open an issue!