Repository navigation
test(search): search-everywhere latency and qmd on local documents - #38
Merged
Merged
Conversation
everywhere/: what a bare word costs once it also looks in transcripts, file names and file text (cli-messaging #1025), at 100k messages, on three builds. local-documents/: qmd 2.8.3 against memo and drive on the owner's notes and Drive documents, 50 known-answer queries. Totals only; queries and per-query ranks stay in cli-private. CLAUDE.md: every measurement of a WireCat CLI lands here, opening with why it was run. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
leemour
added a commit
to WireCatLabs/cli-core
that referenced
this pull request
Oct 13, 2026
It lives at cli-testing performance/search/everywhere/ (WireCatLabs/cli-testing#38), with the results and why it was run. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… had none The first run searched drive's documents by words alone: they were synced without --embed, while search answered by: "words and meaning". After embedding, drive gets 0.68 first and 0.80 in the top 3 in 1.3 s; qmd's best is 0.56 and 0.80 in 11-25 s. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
leemour
added a commit
to WireCatLabs/cli-core
that referenced
this pull request
Oct 13, 2026
…ext (#1025) * fix(speech): make kept voice transcripts searchable, shown and reused Transcripts were already kept in message_transcripts, but nothing else read them: `messages show` did not show them, `content:` search never saw them, and `messages transcribe` heard the message again every time. - A kept transcript becomes the voice attachment's text (extractor `speech:<source>`), so `content:` and the file word index find it. A transcript made before its message was stored is copied when the message arrives; `store reindex` fills in transcripts kept before this change. Agent-written attachment text still wins. - `messages show` and `messages context` read kept transcripts like `messages list` does. - `messages transcribe` answers a kept transcript, and hears again only when a local model other than the one that heard it is asked for. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * wip: a bare search word looks in transcripts, file names and file text Tests green; latency at 100k messages not measured yet, lint not re-checked. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(search): import MatchPlace through store.ts Services reach the store through store.ts only (Biome noRestrictedImports). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * perf(search): find file names through a word index (store version 5) A bare word read every stored file name into JavaScript on each search: at 100k messages and 50k files a search took ~390 ms instead of ~120 ms, even with no match. attachment_name_words is an FTS5 index kept by triggers and filled once on upgrade. The triggers fold й and ё the way normalize() does, since unicode61 folds only Latin accents (owner, NEED-993 A). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * test(search): file names stored before version 5 are found; bench results everywhere-results.md keeps the 100k-message latency of a bare word on three builds: before #1025, names read in JS, and the name index. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * test(search): move the search-everywhere bench to cli-testing It lives at cli-testing performance/search/everywhere/ (WireCatLabs/cli-testing#38), with the results and why it was run. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * docs(search): help names transcript: beside content: and filename:; regenerate the field table The search help #1015 added named content: and filename: only. A bare word now looks in four places, so messages' help says so and lists the four fields; search all's help names transcript: too. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(search): content: no longer claims voice transcripts in the field table Merging main brought back #1010's wording, since the squash of #1010 left the merge base behind it. A merge with the squash commit as its base gives this file exactly; the rest of the tree already matched. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * docs(changelog): keep the search-everywhere and store version 5 entries the merge dropped Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * docs(changelog): restore the search-everywhere, store version 5 and sheet rows entries Merging main dropped them again; #1016 reached main without its entry, so it is added back here. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… limits of both Adds to the local-documents page: a side-by-side of the two pipelines (stemming, vector model, fusion, qmd's query expansion and reranker), why memo and drive put the right document first more often while qmd's hybrid wins on paraphrases, what reranking we already have (a rule reranker in message search, a measured but unshipped cross-encoder) and what is worth taking from qmd. "Not measured" becomes "Limits of this test" with the size, labelling and model caveats. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…wen3 over memo and drive mmarco lifts memo's notes from 0.72 to 0.84 first place and every paraphrase into the top 3 at ~2 s a query in a warm process; on the Drive documents the gain is within noise. Qwen3 (qmd's) is no better and three times slower. Scripts run read-only against the CLIs and keep the owner's data outside the repository. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…f the reranker run Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ded; rerank none Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
The 2026-10-12 field test (tg, memo and drive on a real project) raised two measurements. Both now live here, under the new rule that every measurement of a WireCat CLI lands in cli-testing and opens with why it was run.
What
performance/search/everywhere/: a bare search word now also looks in voice transcripts, file names and file text (feat(search): a bare word looks in transcripts, file names and file text cli-core#1025). At 100k messages, the first build cost ~270 ms per search, because it read every file name. A word index of file names brought this to 15–40 ms over message text alone. The script moves here from that PR.performance/search/local-documents/: qmd 2.8.3 against memo and drive on the owner's notes and Drive documents, with 50 known-answer queries. drive finds only 2 of 15 paraphrases, typos and questions in its top 3; qmd's vectors find 11, and so does memo. qmd takes 10–25 s per one-shot query and needs 2.2 GB of models. Conclusion: give drive memo's search rather than adopt qmd. Totals only; the queries and per-query ranks stay in cli-private.CLAUDE.md: a "Where results live" section.biome.jsonlets the new script print.TOOL-RESEARCH.mdandPROCESS.mdlink both pages.Checks
pnpm lintandpnpm docs:checkpass. No suite orsrc/change.🤖 Generated with Claude Code