09

RAG Reranker

jev primitiveScore per chunk (parallel calls) · Noul (grade)
labelsanswerno_answer
dataset20 labeled rows
tracingLangfuse

Decides what order the retrieved chunks go in, because only the top 3 reach the answer. A word-overlap retriever returns 10 chunks from a 30-chunk help centre, with overview and FAQ pages that repeat the question's words but don't answer it. Jev gives each chunk its own relevance Score (irrelevant, background, directly answers), all 10 calls in parallel, and the chunks are sorted by it. The same model answers from the top 3, and Jev grades it. Measured: nDCG@3 and MRR against gold relevance grades, plus pass rate and cost per passing answer.

With Jev

◐ ▮▮▮ ▦

Without Jev

LLM · CODE
Examples

Calls Jev and every baseline for real, traced in Langfuse. Nothing is saved.