Fix: Clean Residual Edges from VDB During Entity Deletion

[danielaskdd]

Fix: Clean Residual Edges from VDB During Entity Deletion

🐛 Problem

When deleting entities from the knowledge graph, the system was only removing nodes from the graph structure but not cleaning up associated edge data from, causing data inconsistencies between Graph DB and VDB.

• Vector database (relationships_vdb)
• Relation chunks storage (relation_chunks)

This resulted in orphaned edge is cause by manully editing entity name.

✅ Solution

This PR adds comprehensive edge cleanup logic before node deletion:

1. Edge Tracking: Collect all edges connected to entities being deleted into a deduplicated set with normalized representation (sorted tuples for consistency)

2. VDB Cleanup: Delete edge vectors from relationships_vdb using computed relationship IDs for both directions (src→tgt and tgt→src)

3. Storage Cleanup: Remove relation chunk tracking data using make_relation_chunk_key

4. Improved Logging:

   • Fixed edge direction in warning message ({src} <-- {tgt} instead of {tgt} --> {src})
   • Added info log showing number of cleaned residual edges

📝 Changes

• File: lightrag/lightrag.py
• Location: Entity deletion logic in the cleanup process
• Lines Modified: ~30 lines added/modified

🧪 Testing

Tested with entity deletion scenarios to verify:

• All edge data is properly cleaned from VDB
• Relation chunk storage is correctly updated
• No orphaned data remains after deletion
• Log messages provide clear feedback