Skip to content

Improve duplicate-surfacing search recall (in:title + high-recall queries) - #6410

Merged
denelon merged 1 commit into
microsoft:masterfrom
denelon:denelon/dup-surfacing-search-recall
Jul 29, 2026
Merged

Improve duplicate-surfacing search recall (in:title + high-recall queries)#6410
denelon merged 1 commit into
microsoft:masterfrom
denelon:denelon/dup-surfacing-search-recall

Conversation

@denelon

@denelon denelon commented Jul 29, 2026

Copy link
Copy Markdown
Collaborator

📖 Description

The agentic Duplicate Surfacing workflow (#6399 / #6400) reasons about duplicates semantically, but its only retrieval mechanism is GitHub's full-text issue search, whose relevance ranking buries older, sparsely-worded issues. Genuine duplicates that never surface in search can never be judged by the model.

Concrete miss (found by @Trenly, #6406): on #6405 ("Support source type detection") the agent ran jargon-heavy queries (source add --type PreIndexedPackage Rest autodetect, etc.) and no-op'd — but #1477 ("The type of a source should be automatically detected when possible") is a near-verbatim duplicate (same feature and same proposed /information-endpoint mechanism). GitHub full-text search never returned #1477, even for its own title words; only in:title concept searches surface it.

This PR updates the workflow's "How to search" guidance to raise recall:

  • Always run at least two in:title concept searches (2–3 core nouns).
  • Prefer short, high-recall queries; don't pile many exact CLI tokens into one query.
  • Vary noun/verb vocabulary (weak stemming) and scan ~20–30 results per search.

The change is prompt-only — detection thresholds, safe-outputs, permissions, and action pins are unchanged.

Authored with GitHub Copilot assistance.

🔗 References

🔍 Validation

✅ Checklist

📋 Issue Type

  • Bug fix
  • Feature
  • Task
Microsoft Reviewers: Open in CodeFlow

…ries)

The Duplicate Surfacing agent judges semantically but retrieves via GitHub's
full-text issue search, whose relevance ranking buries older, sparsely-worded
issues. On issue microsoft#6405 (Support source type detection) the agent's jargon-heavy
queries missed the true duplicate microsoft#1477 (The type of a source should be
automatically detected when possible) — GitHub full-text search never returned
it, even for its own title words.

Update the 'How to search' guidance to:
- Always run at least two in:title concept searches (2-3 core nouns).
- Prefer short, high-recall queries; avoid piling exact CLI tokens into one
  query, which shrinks the result set.
- Vary noun/verb vocabulary (weak stemming) and scan ~20-30 results per search.

Validated via 'gh aw trial' replaying microsoft#6405 against the live corpus: the agent
now surfaces microsoft#1477 with HIGH confidence.

Reported by Trenly on microsoft#6406.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 3701d64a-96ba-42f1-be59-b4e7d2a2820a
@denelon
denelon marked this pull request as ready for review July 29, 2026 18:40
@denelon
denelon requested a review from a team as a code owner July 29, 2026 18:40
@denelon
denelon merged commit cba7c0d into microsoft:master Jul 29, 2026
4 checks passed
@denelon
denelon deleted the denelon/dup-surfacing-search-recall branch July 29, 2026 18:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Improve duplicate-surfacing search recall (in:title + high-recall queries)

3 participants