improvement(search): rank live results across sources by the query terms they hold

[waleedlatif1]

Summary

• Live search fused every account's results by 1 / (RRF_K + rank) within that account, so each account's first result tied with every other account's first result and the stable sort kept account order: the merged list was a round-robin of Drive, Gmail, Slack, Jira, Confluence regardless of the query. The topK cut then fell before any cross-source comparison, so a page whose title is the query sat behind a file-name-only Drive hit and an unrelated Jira ticket, and a source's 5th-best match never surfaced even when it beat every other source's best
• A plain query now orders the merged results by queryRelevance: the distinct (stemmed) query terms the title and preview hold, then the terms the title holds, then the share of the title's words that are query terms; fused similarity breaks the remaining ties
• Date-sorted, queryless and native-query searches keep fusion order (native queries carry provider operators, and were not measured)
• Applies to the Search tab and the Search MCP / assistant tools alike; no new provider requests

Type of Change

• Improvement

Testing

• application.test.ts: new case where a later result holding the query (and a title named for it) must outrank another account's first result; it fails on the pre-change sort. application.test.ts + term-fallback pass (73)
• Offline replay on a large production tenant (read-only): each source's own ranked list captured at topK 20 (the Search tab and MCP default; reconstructs the live merged order 19/20 positions), every pooled document labeled blind by two independent judges, product queryRelevance imported directly:
  • 10 keyword queries from the customer's own Search-vs-competitor check (dev): nDCG@10 0.42 → 0.77 (Opus) / 0.38 → 0.73 (Sonnet); top result relevant 1 → 7 of 10; irrelevant top result 3 → 0
  • 66 held-out queries issued by a Claude client during benchmark sessions, judged against each question's expected facts: nDCG@10 0.29 → 0.65 / 0.28 → 0.65, better on 65/66 and 66/66 queries, worse on 1/0; expected-fact documents in the top 10 0.28 → 0.62, in the top 20 0.48 → 0.81; irrelevant top result 89% → 21%
  • Variants rejected on the same data: coverage only, title first, a weighted sum, pool-IDF weighting (each lower on held-out nDCG and recall, some with 5–13 losses)

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Oct 16, 2026 7:04pm UTC

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Plain live searches now compare results across accounts by the query terms in their titles and previews before applying topK.

• Plain searches rank results across accounts by query matches.

No actionable issues were found. This review used code inspection; tests were not run.

_{Reviews (1) · Last reviewed commit: "improvement(search): rank live results a..." · Reviewed by Greptile}

[cubic-dev-ai]

All reported issues were addressed across 4 files

<sub>Reply with feedback, questions, or to request a fix.

Turn on auto-fix | Re-trigger cubic</sub>