Priority of search results when using searchActorTypeahed

[myConsciousness]

Hi,

When we search for @yui.syui.ai, we first enter @yui. as follows, but @yui.syui.ai is sorted as the 6th. As a matter of priority, it's possible that @yui.bsky.social, the handle is very similar to @yui.syui.ai, will be sorted as the first one, but in this example @yui.bsky.social is appear as the fifth one.

Perhaps the results retrieved by displayName are preferred over handle.

atproto/packages/bsky/src/services/util/search.ts

Lines 46 to 135 in 4619e2b

	export const getUserSearchQuerySimple = (
	db: Database,
	opts: {
	term: string
	limit: number
	},
	) => {
	const { ref } = db.db.dynamic
	const { term, limit } = opts
	// Matching user accounts based on handle
	const accountsQb = getMatchingAccountsQb(db, { term })
	.orderBy('distance', 'asc')
	.limit(limit)
	// Matching profiles based on display name
	const profilesQb = getMatchingProfilesQb(db, { term })
	.orderBy('distance', 'asc')
	.limit(limit)
	// Combine and paginate result set
	return paginate(combineAccountsAndProfilesQb(db, accountsQb, profilesQb), {
	limit,
	direction: 'asc',
	keyset: new SearchKeyset(ref('distance'), ref('actor.did')),
	})
	}
	// Matching user accounts based on handle
	const getMatchingAccountsQb = (
	db: Database,
	opts: { term: string; includeSoftDeleted?: boolean },
	) => {
	const { ref } = db.db.dynamic
	const { term, includeSoftDeleted } = opts
	const distanceAccount = distance(term, ref('handle'))
	return db.db
	.selectFrom('actor')
	.if(!includeSoftDeleted, (qb) =>
	qb.where(notSoftDeletedClause(ref('actor'))),
	)
	.where('actor.handle', 'is not', null)
	.where(similar(term, ref('handle'))) // Coarse filter engaging trigram index
	.where(distanceAccount, '<', getMatchThreshold(term)) // Refines results from trigram index
	.select(['actor.did as did', distanceAccount.as('distance')])
	}
	// Matching profiles based on display name
	const getMatchingProfilesQb = (
	db: Database,
	opts: { term: string; includeSoftDeleted?: boolean },
	) => {
	const { ref } = db.db.dynamic
	const { term, includeSoftDeleted } = opts
	const distanceProfile = distance(term, ref('displayName'))
	return db.db
	.selectFrom('profile')
	.innerJoin('actor', 'actor.did', 'profile.creator')
	.if(!includeSoftDeleted, (qb) =>
	qb.where(notSoftDeletedClause(ref('actor'))),
	)
	.where('actor.handle', 'is not', null)
	.where(similar(term, ref('displayName'))) // Coarse filter engaging trigram index
	.where(distanceProfile, '<', getMatchThreshold(term)) // Refines results from trigram index
	.select(['profile.creator as did', distanceProfile.as('distance')])
	}
	// Combine profile and account result sets
	const combineAccountsAndProfilesQb = (
	db: Database,
	accountsQb: AnyQb,
	profilesQb: AnyQb,
	) => {
	// Combine user account and profile results, taking best matches from each
	const emptyQb = db.db
	.selectFrom('actor')
	.where(sql`1 = 0`)
	.select([sql.literal('').as('did'), sql<number>`0`.as('distance')])
	const resultsQb = db.db
	.selectFrom(
	emptyQb
	.unionAll(sql`${accountsQb}`) // The sql`` is adding parens
	.unionAll(sql`${profilesQb}`)
	.as('accounts_and_profiles'),
	)
	.selectAll()
	.distinctOn('did') // Per did, take whichever of account and profile distance is best
	.orderBy('did')
	.orderBy('distance')
	return db.db
	.selectFrom(resultsQb.as('results'))
	.innerJoin('actor', 'actor.did', 'results.did')
	}

[mozzius]

Another example:

[henoya]

It seems that commit b556912 on 2023

-08-08 fixed the search similarity formula related to this Issue, and the issue poster's issue has been fixed, but I found another problem example.

When I try to search for the "@mmm.bsky.social" account in a post or in the search input, typing "@mmm" results in other accounts such as "@Mmmm" coming up on top. This may be unavoidable, but if I type "@mmm.bsky" further, the "@mmm.bsky.social" account comes up as the 6th account, and it appears lower in the list of candidates in the PC Web version, but not in the list of 5 candidates in the mobile version.

If there is a clear separation between words, it would be better to prioritize similarity on a word-by-word basis.

[syui]

User completion was working until a week ago.
However, it appears that user completion is not working again.

[bnewbold]

@syui do you have a more specific example of something broken right now?

We did deploy some pretty large changes to profile ("actor") search recently; the corresponding post search changes haven't gone live yet. I think that most of the issues raised above are at least partially resolved, based on my spot checks right now.

We do know that a very small fraction of accounts are currently not indexed in the profile index, if they do not have a "profile" record in their repository (even an empty record). It is also possible that a handful (dozens?) of accounts have failed to re-index.

[syui]

display name : ai
handle : yui.syui.ai

These may be a matter of priorities.