Summary

• Improve _extract_and_filter_label to handle varying LLM output formats
• Strip markdown code blocks before JSON extraction
• Support both {"vertices":[...], "edges":[...]} (object) and flat array formats
• Auto-convert flat arrays to the expected object structure

Problem

When using reasoning models (e.g., DeepSeek V4) for graph extraction, the LLM may return:

1. JSON wrapped in markdown code blocks (\``json ... ```), which breaks the greedy regex ({.*})`
2. A flat array [vertex, edge, ...] instead of the expected object {"vertices": [...], "edges": [...]}

Both cases cause json.JSONDecodeError and result in empty extraction output even though the LLM correctly identified entities and relationships.

Solution

• Strip markdown code fences (\``json/````) before regex matching
• Update regex to match both objects ({...}) and arrays ([...])
• When a flat array is detected, partition items by type field into vertices and edges

Test plan

• Test with OpenAI models (existing behavior should be preserved)
• Test with DeepSeek models (markdown-wrapped array format)
• Test with Ollama models
• Verify both object and array formats are handled correctly

🤖 Generated with Claude Code