Measured on the corpus (2026-10-01): the query **`how do I fix my car engine oil change`** returned **2
Measured on the corpus (2026-10-01): the query how do I fix my car engine oil change returned 2 results. The query best pizza recipe with mozzarella returned none. Both are outside what this corpus knows about, and only one of them was answered anyway.
That asymmetry is the whole problem. A search that returns nothing is legible: the caller learns "there is nothing here". A search that returns two plausible-looking lessons for a question about engine oil is not — and the caller is an agent that has been told to treat retrieved content as evidence. A wrong hit is worse than no hit, because it is consumed silently.
Two different questions get collapsed into one similarity number:
Lexical scoring answers the first and only approximates the second. Rare terms make this worse in a specific way: a query word that is absent from the whole corpus has maximal idf, so it inflates the score of any document that matches any other word — and if the denominator and numerator of a relevance floor are computed over different term sets (say, the user's words versus the alias-expanded ones), an expansion term can supply mass the denominator never contained. That is how a query with no coverage at all clears a 0.55 coverage floor at 0.677.
The second half of the cause is design rather than maths: teams treat "returned something" as success, so the no-match path is never built. There is no honest answer available for the case that is actually common — the corpus does not cover this topic yet.
no_match a first-class answer, not an empty array with a 200. Say that nothing matched, and carry the next step with it: the intake path, so the gap can be filed instead of guessed at. An agent that knows the corpus is empty on a topic will say so; an agent handed two weak hits will cite them.serve and mcp are legitimate queries here, and a blocklist tuned on pizza will refuse them next month.node --test workers/search-low-relevance-queries.test.mjsSix boundary queries must come back no_match (pizza, engine oil, sourdough, keyboard mashing, 0x0a, and a sentence pulled from the corpus's own chatter), and one covered query — pip install timeout corporate proxy — must still find pip-install-proxy-timeout. Both halves in one file, because either half alone can be satisfied by breaking the other.
The historical detail is worth keeping in the test: the engine-oil query returned 2 results before the floor work, so a future failure there is a regression with a known shape, not a new mystery.
When a search returns results, check why they matched before using them — the query's own words, the domain, the rank — and treat "nothing matched" as information rather than a failure to work around. When you report a retrieval result, quote the query and the ids you got; if a query returned hits that do not look like the topic, that is a gap to file, and the boundary queries above are the shape of the evidence that gets it fixed.