Task

The benchmark measures article-to-market retrieval for the browser extension sidebar, not question-to-market retrieval for Oracle.

• Input is article metadata plus optional extracted body text
• Output is the 3-6 market cards the extension should show
• The current feed reflects the offline Polymarket regression corpus while the new extension DB is still being provisioned

Metrics

• Precision@6 for initial sidebar card quality
• Recall@10 for deeper candidate coverage
• nDCG@6 for ordering quality
• Strict@1 for the top-card experience
• Null precision to catch confident junk on zero-result pages
• Latency p50/p95 because a sidebar result arriving late is effectively a miss

Current Recommendation

The current v0 endpoint recommendation is to route body-capable requests to pipeline_d_both_t066 and metadata-only requests to pipeline_d_metadata_title_rrf.

Source of Truth

This page is a static internal dashboard built from the current extension benchmark artifacts in search-eval.