fix: normalize Milvus similarity scores for threshold filtering

[ejrtks1020]

🐛 Fix: Normalize Milvus similarity scores for SimilarityThresholdRetriever

Summary

Added score normalization logic to Milvus vector store to ensure SimilarityThresholdRetriever works correctly across different metric types.

Changes

• Added normalization logic for L2, IP, and COSINE metric types
• L2 distances converted to similarity scores using 1/(1+distance)
• IP and COSINE scores normalized from [-1,1] to [0,1] range

Problem Solved

• SimilarityThresholdRetriever was not working properly with Milvus due to unnormalized scores
• Different metric types returned scores in different ranges, causing inconsistent threshold filtering
• Raw distance/similarity values were passed directly without proper normalization

Testing

• Tested with L2, IP, and COSINE metric types
• Verified SimilarityThresholdRetriever now works consistently across all metric types
• Confirmed scores are properly normalized to 0-1 range

Files Modified

• packages/components/nodes/vectorstores/Milvus/Milvus.ts

related issue #4879

[HenryHengZJ]

thank you!