Milvus Hybrid Retrieval: Weighted Fusion Scores vs. Similarity Thresholds
Β· 3 min read
Debugging hybrid retrieval scoring in a RAG knowledge-base project β the full troubleshooting trail below.
TL;DRβ
Milvus hybrid retrieval with weighted fusion scores as 0.7 * dense_score + 0.3 * sparse_score tops out around 0.7 in theory β and lower in practice. Filtering with min_similarity=0.7 removes nearly everything. Fix: drop the threshold to 0.3, or adapt it to the fusion strategy dynamically.