08

RAG Context Filter

jev primitiveNoul ×3 per chunk (parallel calls) · Noul (grade)
labelsanswerno_answer
dataset20 labeled rows
tracingLangfuse

Decides which retrieved chunks the answer may use. A plain word-overlap retriever returns the top 6 chunks from a 20-chunk help centre, noisy on purpose: a stale 2019 policy, a blog post, an injected instruction, internal salaries. Jev judges each chunk on its own (useful, injection, sensitive: three Nouls), all in parallel, and a policy keeps at most 4. The same model then answers from what was kept, and Jev grades every answer against a rubric. Measured: pass rate, cost per passing answer, answer vs NO_ANSWER, and retrieval precision and recall.

With Jev

◐ ▮▮▮ ▦

Without Jev

LLM · CODE
Examples

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