Repository navigation
Replies: 2 comments
|
An independent experiment on Postgres, hybrid search (vectors on), gbrain 0.60.376. On 38 synthetic Japanese questions with expected pages fixed in advance, Recall@5:
In this sample, every sentence-form query without a date filter reported Caveat: the questions were generated from page text, so the absolute numbers favor retrieval. In this synthetic sample, the gap is worth investigating; it does not establish production retrieval accuracy. This resembles the query-form gap reported in #6888, though the setups differ (PGLite keyword-only there) and do not establish whether the effect is engine-specific. |
|
A small query-side experiment, using synthetic data. Change (local branch on master Examples: 「バックアップはどうすればいい?」 → Measurement: a fresh PGLite brain with embeddings off (keyword-only), 24 fictional Japanese note pages, and 24 sentence-form questions with one expected page each.
Across the 28 hits (both commands), the expected page ranked first in 26; no question returned only wrong pages. The CJK and keyword test files (16 files, 190 tests) and typecheck passed. Limits: the other 10 questions still returned nothing. The keyword-only path requires every term, and for each of the 10 I checked that at least one segmented term does not occur in the expected page: question words (for example 何時, 原因, 必要) or verb forms that differ from the page. I have not measured the hybrid arm, whose any-term fallback might recover some of them, nor the effect on the CJK deadline on a large corpus. The question set is small and written by me, so the numbers show the direction, not production accuracy. |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #6888, opened by @Jumbo0714 on 2026-10-17. Feature requests now live in Discussions > Ideas.
Environment
What happens on the keyword path
src/core/cjk.ts,src/core/search/cjk-keyword-sql.ts): split on whitespace only, every piece must appear verbatim (ILIKE, AND, unindexed).cjk.ts.Measured with keyword search only (0.60.17, 23 Japanese questions with expected pages fixed in advance)
With bge-m3 embeddings the same natural questions reach 23 / 23, so the gap is specific to the keyword arm.
Related
Proposal
Intl.Segmenter('ja', { granularity: 'word' })(built into Bun and Node, no dependency; it splits 「AIハーネスは開発中のたたき台として継続する」 into sensible words) and character bigrams as extra projections, use the same tokenization for queries, and keep the current ILIKE path as a fallback.Happy to test a branch against the same question set on Windows/PGLite.
All reactions