GET /xrpc/tech.waow.typeahead.searchActors typeahead.waow.tech

Multi-word queries are strictly conjunctive and miss near matches #38

Open zzstoatzz.io
Open opened by zzstoatzz.io

On 2026-11-10 I sent 34 queries once each to this service, to tern and to Bluesky's public typeahead, at about one request a second per service, and checked all 384 returned accounts against app.bsky.actor.getProfiles.

What agreed #

  • All three ranked the exact handle first for 11 handles, 8 of them on independent PDSes.
  • No service returned a handle that differed from the appview's current one.

Where this service was wrong #

paul frazee returned one account and not pfrazee.com. Bluesky returned it third. nate kohari worked. bench.py already lists paul frazee among its display-name queries.

Correction, same day: I first wrote that the account's display name is the query. It is P(aul) Frazee. The cause and the two possible changes are in the first comment; the engine is doing what it was designed to do.

What the comparison did not cover #

Every returned account exists on Bluesky, so the set never reached the cases where the designs differ. A second round should seed cases from this service's own data:

  • accounts with no Bluesky profile (for example Tangled-only or place.birds.* users)
  • handles changed in the last hour, to compare how quickly each service reflects them
  • accounts deleted, deactivated or taken down in the last week
  • accounts carrying moderation labels this service hides

Keep the same manners: one pass, about a request a second, a User-Agent that names a contact.

Other observations #

  • tern returned one account labelled spam; this service and Bluesky returned none.
  • tern ranks a handle whose first label equals the query above much larger accounts (news.sky.com before bloomberg.com for news), which is its stated preference.
  • Median response from one location: 244 ms here, 272 ms tern, 138 ms Bluesky. One location and one pass; not a latency claim.

Diagnosed on 2026-11-10 with ten polite live requests and the normalizer built locally. No code was changed.

Cause

  • pfrazee.com's display name is P(aul) Frazee, in Bluesky's profile and in this index.
  • Parentheses are separators (services/src/index/normalize.zig:74-78), so the name normalizes to p aul frazee and the actor has no key for paul.
  • A multi-token query requires every token (services/tantivy-normalizer/adapter.zig:186,193-198; services/tantivy/src/engine.rs:870-885), and in that mode the handle and domain alternatives are not tried.
  • So paul frazee needs both tokens and finds only an account whose name contains both. frazee alone returns pfrazee.com first.

Bluesky is not conjunctive: for the same query it also returns Paul Krugman and other Pauls with no frazee, and ranks pfrazee.com on a partial match plus popularity.

Two possible changes, both policy decisions

  1. Also index a joined token, so P(aul) emits paul and O'Brien emits obrien. Adds keys only, so existing canonical results should hold, but it bumps the normalizer version and needs a full index rebuild and a slot image release.
  2. Fall back to partial-token matches when the conjunctive result is short of the limit. Query-side only, no reindex, but it changes every sparse multi-word query: nate kohari would gain a tail of other Nates.

A smaller related gap: paulfrazee.bsky.social, Bluesky's second result, is missed because multi-token queries never try the tokens joined as a handle prefix.

Not verified: whether the Worker's result validation would accept partial matches under option 2.

Sign up or Sign in to add to the discussion.