Database / Weaviate Vector database Interview questions
What is the difference between rankedFusion and relativeScoreFusion in hybrid search?
Both are algorithms Weaviate can use to combine the separate vector-search and BM25-search result lists into one fused ranking, but they combine those lists using different mathematical approaches.
| rankedFusion | relativeScoreFusion |
| Combines results based on each item's rank position in each list. | Combines results based on normalized, relative scores from each list. |
| Less sensitive to how far apart raw scores are between top and lower results. | Preserves more information about how much better one result is than another. |
| More robust when the two scoring systems aren't well calibrated against each other. | Can produce more nuanced rankings when score magnitudes are meaningful and comparable. |
rankedFusion, similar in spirit to Reciprocal Rank Fusion used elsewhere in the industry, tends to be the safer default when there's no strong reason to trust that BM25 scores and vector similarity scores are on genuinely comparable scales. relativeScoreFusion can produce better results when that comparability roughly holds, since it uses more of the actual signal strength each list carries rather than reducing everything to rank position alone.
rankedFusion combines results primarily based on:
relativeScoreFusion, compared to rankedFusion, is more likely to:
More Related questions...
