A postinstall patch rewrites `dist/store.js` after npm unpacks it, and Open Walnut had 15 of them pinned to a single version of QMD 2.1.0 [12]. That kind of edit changes what a function returns on the way out, not which tokens went into the index. FTS5 was configured with `porter unicode61` [20], so the indexing decision was made in the engine core long before any wrapper saw the query. Of the three causes the team traces the August miss to, the tokenizer, the AND-joined query and the scoring were all inside the library, while the patches and roughly 3,700 lines of integration code across 18 files sat outside it [21][12].
Those 3,700 lines are 4.6 times the roughly 800 lines of the first engine the team threw away [2][8]. Several of the 18 files reached into the library's private `internal.*` API [12].
The latency profile says who the library was built to serve. `/api/search` averaged 4,034 ms in production against a hard 300 ms prefetch budget, which is 13.4 times over [14][5][1]. The cross-encoder reranker took 9.29 seconds at p50 to rerank 40 candidates, about 232 ms per candidate [15][3], and bought 10 points of observed Recall@10, from 81.7% to 91.7% [15][6]. Ten points is a real gain, but it is not available to anything a person is watching, so it was disabled on every interactive path [15].
One reason for leaving the first engine was that Ollama was a second process the user had to install and keep running [9]. Synchronous, seconds-long index writes then forced a forked child process to keep the web server's event loop alive [17]. The same process problem recurred.
Whether any of these numbers is about your workload depends on the corpus. QMD was built for markdown knowledge bases, mostly in one language [18]. This index holds about 12,000 documents, of which 3,593 are session transcripts, and the session store alone is 388 MB of the 663 MB on disk, 58.5% of the total [6][13][4]. About 31% of task titles mix Chinese and English, and notes carry identifiers like `CR-291543784` [7]. Point the same library at single-language prose in files and the tokenizer complaint goes away, and most of the patch count with it.
The eleven days is the team's own accounting, and it is cheap for a reason it states plainly: once tokenization moved into their code, the rest of the engine turned out to be things already in their stack [4][23]. A team that genuinely wanted the reranker and the query-expansion model, rather than shipping a 3.8 GB model cache with more than half of it never enabled [16], would be rewriting considerably more than a tokenizer.
The July embedding study is the weakest link in the chain. It ran on an isolated copy of 3,549 tasks, through QMD's hybrid path, and Qwen3 won and is still the default in the engine that replaced that path [22]. The comparison table is not in the write-up available here, so the default rests on a measurement taken under scoring that no longer runs.
The stage that decided what existed to be found was the one stage this library did not let anyone replace, and that is the standard a search library should be judged by [3].
