Skip to content

[eval only, do not merge] Reword search_issues for natural-language semantic queries#2937

Closed
kelsey-myers wants to merge 2 commits into
mainfrom
km/eval-search-issues-semantic-wording
Closed

[eval only, do not merge] Reword search_issues for natural-language semantic queries#2937
kelsey-myers wants to merge 2 commits into
mainfrom
km/eval-search-issues-semantic-wording

Conversation

@kelsey-myers

Copy link
Copy Markdown
Contributor

Do not merge — eval scaffold

Draft branch for the copilot-offline-eval MCP tool-selection harness. It exists so the eval can containerise this branch (github-mcp-server-branch input) and measure query formation against the current lexical wording. Not intended to land.

What it changes

Rewrites the search_issues tool description and query param description to steer models toward natural-language queries instead of qualifier/boolean syntax:

  • Tool: …using natural-language semantic matching. Write the query as a natural-language description of what you're looking for… Avoid boolean OR operators, which fall back to lexical keyword search.
  • query param: Write it as a natural-language description… rather than exact keywords or boolean operators.

This matches the end-state of the issue-search experiment (github/plan-track-agentic-toolkit#415): once search_issues defaults to semantic, the description should stop advertising lexical syntax.

Why

The experiment showed models over-produce boolean/qualifier queries that trip the monolith's semantic→lexical fallback. Before committing to the remote-side wording, we want offline eval evidence that the reworded description actually shifts query formation (QUERY_AGENT_RATING / QUERY_WORD_MATCH) without regressing legitimate qualifier queries (guard cases in the benchmark PR).

Validation

  • go build ./cmd/github-mcp-server
  • go test ./pkg/github/ -run Test_SearchIssues
  • gofmt clean ✅
  • Toolsnaps regenerated (search_issues.snap, search_issues_ff_fields_param.snap)

Draft branch for the copilot-offline-eval MCP tool-selection harness. Rewrites
the search_issues tool + query param descriptions to steer models toward
natural-language queries instead of qualifier/boolean syntax, matching the
end-state of the issue-search experiment (plan-track-agentic-toolkit#415).

Not intended to merge: it exists so the eval can build a container from this
branch and measure query-formation against the current lexical wording.

Toolsnaps regenerated.
Realign the eval stand-in to the exact strings shipped by the experiment PR
(github-mcp-server-remote#911) semantic variant, rather than the stronger
end-state wording, so the eval measures what the experiment actually ships:

- tool desc: 'Best for conceptual or paraphrased queries...' (not the imperative
  'Write the query as a natural-language description...')
- query param: 'Write it as natural language for semantic and hybrid variants.'
  (drops 'rather than exact keywords or boolean operators')

Toolsnaps regenerated.
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.

1 participant