mirror of
https://github.com/lancedb/lancedb.git
synced 2026-10-15 00:02:26 +00:00
brekenandbreken-ai
c119f0c8e5
## What's wrong
`RRFReranker.rerank_multivector` concatenates every vector result list
and passes the concatenation to `rerank_hybrid` as a single list.
`rerank_hybrid` then scores each row by its position in that
concatenation, so a row's rank depends on how many rows came before it
in *earlier* lists.
With two searches of 3 rows each (`K=60`):
| row | rank in its own list | score today | RRF score |
|---|---|---|---|
| 1 (top of list 1) | 1 | 1/61 | 1/61 |
| 3 (third of list 1) | 3 | 1/63 | 1/63 |
| 4 (top of list 2) | 1 | **1/64** | 1/61 |
| 5 (second of list 2) | 2 | **1/65** | 1/62 |
So the best hit of the second search ranks below the worst hit of the
first, and the order of the lists passed in decides the fused ranking.
`MRRReranker.rerank_multivector` already scores each list on its own
("MRR semantics require treating each vector result as a separate
ranking system"). RRF is defined the same way.
## Fix
Score each result list with its own ranks (`1 / (rank + K)`, summed
across lists for rows found by more than one), then dedupe the
concatenation by `_rowid`, attach `_relevance_score`, and sort. The
`return_score` handling is unchanged.
## Tests
`test_rrf_multivector_ranks_each_result_list` in
`python/python/tests/test_rerankers.py`:
- On `main`, it fails with `assert 0.015625 == 0.01639344262295082` (the
top hit of list 2 got `1/64` instead of `1/61`).
- With the fix, it passes, and so do the other tests in
`test_rerankers.py` (16 passed, 12 skipped for missing API keys or
models).
- `ruff format --check` and `ruff check` are clean.
The fix was found and written by an AI agent (Breken) and checked
against the tests above.
Co-authored-by: breken-ai <312387581+breken-ai@users.noreply.github.com>