Evals · conare-rerank-1

Several conditions at once

Find the one document that meets every listed condition among near-copies that each break one. 474 searches over books, legal cases, medical records, movies and people.

conare-rerank-1

2.5%

% ranked wrong, lower is better · #2 of 6, level with Claude Opus 5.5

  1. 1Claude Opus 5.5Level with conare-rerank-1 within the 95% interval1.1%
  2. 2conare-rerank-12.5%
  3. 3GPT-6 SolLevel within the 95% interval3.8%
  4. 4Mixedbread mxbai-rerank-v3.116.7%
  5. 5Contextual rerank-v250.0%
  6. 6zerank-263.5%

% ranked wrong, lower is better · read 14 October 2026

conare-rerank-1 was trained on examples built the way this set is built.

How it was measured

MultiConIR, 5 domains. A search is wrong when the matching document is not ranked first. The LLMs were given each document's first 2,000 characters, so only the 474 of 497 searches whose deciding fact falls inside them are counted, for every model. Run on 27 September 2026, recounted on 14 October 2026.

See it on your own data.

Book a demo