JingsongLi commented on PR #10163:
URL: https://github.com/apache/paimon/pull/10163#issuecomment-5831992175
Closing this PR under the end-to-end CLI-value criterion for this review
pass. The new command has no successful command-level search test, and the
default `--limit 10` output loses the defining nearest-neighbor ranking:
`execute_local()` carries scored top-K row IDs, but `cmd_table_vector_search`
passes the result into a normal table scan and prints that scan's DataFrame
without scores or a score-based reorder. Rows can therefore appear in
row-ID/file order even when the closest match has a larger row ID. The earlier
review reported `bitmap_order=[2,10]` versus `score_order=[10,2]`; the current
head still has this path, and the author has identified the scan-layer ordering
work as a follow-up. With no score shown, a user cannot tell which returned row
is nearest.
I also reproduced two CLI input/output gaps on this head:
`_parse_query_vector('1,,2')` silently returns `[1.0, 2.0]` rather than
rejecting the missing element, and
`json.dumps(_records_for_json(pd.DataFrame({'ts':
[pd.Timestamp('2026-01-01')]})))` raises `TypeError: Object of type Timestamp
is not JSON serializable`. The NumPy embedding conversion fixes one JSON case
but not ordinary projected timestamp columns.
Verification: all 54 CLI table tests passed locally, and vector scoring
tests passed (10 tests plus 35 subtests); the head's Python/Native CI is green.
A broader local vector-table run reached an unchanged optional Vortex case and
segfaulted in the local Python 3.13/PyArrow 24 environment, so it did not
provide a complete local integration result. Please bring the score-preserving
search/read path and at least one real successful CLI test that asserts
nearest-first order and JSON output with typical projected types together in a
new PR. The CLI can then be reviewed as a usable production workflow.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
