perf: eliminate N+1 metadata queries in vector search (searchL1Vector / searchL0Vector)

[yuanrengu]

Summary

Replace per-row metadata lookups with a single batch query in searchL1Vector and searchL0Vector.

Previously, after vec0 returned KNN results, each record_id was looked up individually via stmtGetMeta.get(record_id) — an N+1 pattern that issues up to topK + buffer (default 30) separate SQLite round-trips per vector search.

Now, all valid record_ids are collected and fetched in one SELECT ... WHERE record_id IN (?, ...) query via new batchGetL1Meta / batchGetL0Meta helpers, then resolved from a Map in O(1).

Changes

• batchGetL1Meta(recordIds) — new private method, batch-fetches L1 metadata, returns Map<string, meta>
• batchGetL0Meta(recordIds) — new private method, batch-fetches L0 metadata, returns Map<string, meta>
• searchL1Vector — refactored to use batch lookup instead of N individual stmtGetMeta.get() calls
• searchL0Vector — refactored to use batch lookup instead of N individual stmtL0GetMeta.get() calls

Behavior

No functional change. The zero-vector filtering, orphan detection, result ordering, and fault-tolerant behavior are all preserved:

Aspect	Before	After
Metadata queries per search	N (up to 30)	1
Zero-vector filtering	continue in loop	filter before loop
Orphan handling	warn + skip	warn + skip (same)
Result ordering	vec0 distance order	vec0 distance order (same)
Error handling	catch → empty array	catch → empty Map → empty array (same)

Test plan

• Build passes (npx tsdown)
• Verify vector search returns correct results with existing data
• Verify orphan detection still logs warnings for records with vectors but no metadata
• Verify degraded mode (sqlite-vec unavailable) still returns empty results gracefully

[Maxwell-Code07]

Hi @yuanrengu, thanks for the PR! 👍

The N+1 query optimization in vector search is an interesting area to look at. We'll review the implementation details — including edge cases like empty recordIds and SQLite parameter limits — and get back to you shortly.

Thanks for the contribution! 🙏

[yuanrengu]

@Maxwell-Code07 Thanks for taking a look! Let me know if any changes are needed.