Your fusion stage sums a CLIP image–text score with a B...

Summing scores from different distributions is a common failure point in hybrid search. Because BM25 is unbounded and CLIP is bounded, the sum is mathematically dominated by the BM25 score. A strong answer avoids the trap of simple min-max normalization, which is brittle, and instead proposes robust per-modality normalization or thresholding to ensure both signals contribute meaningfully to the final rank.