chops-search
When it breaks, it says so
Partial loading creates a failure mode most search libraries don't have: what if a row you need isn't loaded? Offline, a strict CSP, a host that ignores range requests. The design principle throughout is that degraded and honest beats complete and wrong.
The wrong answer, and the right one
The tempting implementation averages the rows you have. That produces a
shrunken, wrong embedding that returns plausible-looking garbage, which is
worse than returning nothing, because nobody notices. So embed() returns
nothing instead, search degrades to keyword-only, and it reports that it
did: the status line tells the user semantic matching is off rather than
letting search silently get dumber.
The degradation ladder
| Condition | Behaviour |
|---|---|
| Row fetch fails mid-query | Keyword-only for that query, reported in the status line |
| Host ignores range requests (200 instead of 206) | The whole file is ingested once; eager loading, slower but correct |
Wasm blocked (CSP without 'wasm-unsafe-eval', fetch failure) | Search reports unavailable rather than rendering a dead input |
| Stale page script with fresh artifacts | "Search unavailable", never wrong results; the runtime cache policy bounds this to minutes |
The 200-tolerance is why zola serve works at all: dev servers commonly
don't implement ranges, and a range-hostile host should cost bandwidth, not
correctness. It's also why you should test range behaviour against a real
deploy; locally you're always on the eager path.
Trust, but verify the trusted parts
The same suspicion applies to inputs the browser can't control. A
Content-Range header is checked against what was asked for rather than
believed outright, and artifacts are addressed by content hash from a
revalidating manifest, so a half-updated CDN can't pair a new index with old
rows.
Degradation paths that are designed rather than discovered in an issue tracker have a second benefit: they're testable. The eval harness drives the real plan, range-fetch, ingest, and search path, so a bug in the loading logic shows up as a recall drop instead of hiding behind a fallback.