Disk-resident indexes drop selected docs when the ID selector is smaller than k (no exact-match path)

[chishui]

What is the bug?

detail::should_run_exact_match is called only from nsparse/seismic_index.cpp:171 and
seismic_scalar_quantized_index.cpp:218. No disk_seismic* file calls it, so an enumerable selector
of size <= k takes the block-budget path and most selected docs are never candidates.

How can one reproduce the bug?

2000 docs, dim 1024, 30 nnz/doc, k=10, 20 queries,
8bit|vmin=0|vmax=1|lambda=300|beta=200|alpha=0.4, selector = the query's exact top-n.
Members returned, memory and mmap alike:

selector	seismic_sq	disk_seismic_sq
1	20/20	13/20
10	200/200	113/200

test_disk_seismic_sq_index.py builds lambda=25|beta=4, where disk returns 5/5 — hence no coverage.
disk_seismic_contract.py:150 asserts only that non-members are excluded.

What is the expected behavior?

Selector <= k: return every member, as seismic_sq does. Dispatch to detail::ExactMatcher.

Do you have any additional context?

In OpenSearch neural-search a per_block field filtered to 5 docs with k=10 returns 0 hits.

[zirui-song-18]

Thanks for pointing out of that! This bug is due to the specific structure of DiskSeismic, where vectors will only be stored in an inline format in the posting lists. There is no forward index like seismic index.

To fix this bug, I will propose a PR to introduce doc id mapping in the DiskSeismic structure. So next time when you run exact match, it can fetch the correct document vector and return the exact score.

[zirui-song-18]

Close this issue now as #52 has been merged