JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Retrieval walkthrough

Re-ranking

Builds BM25 shortlists for 40 CLERC legal queries, then asks Jev about each query–passage pair to move answer-bearing passages toward the top of the search results.

Original by TypeSafe AIRetrievalTypeSafe cookbookSource reviewed

Before you dive in

What you’ll find in the original

  1. Retrieve a cheap shortlist before calling a semantic reranker.
  2. Compare top-1 and top-10 retrieval, not only a few convincing examples.
  3. Budget for one decision per query–candidate pair when scaling the shortlist.
Worth knowing

In this 40-query example, top-1 rose from 5% to 18% and top-10 from 38% to 62%; neither is a universal accuracy claim.