RAG passage gate

Who: RAG / search engineers. Steps: Retrieve top-k → per passage Nouls (relevant / usable evidence / contradicts premise / injection) → route() in code → build prompt with accepted vs conflict blocks → LLM answers. Expected effect: drop injections/off-topic; surface premise conflicts; cookbook shows reshuffle vs similarity rank alone (cookbook-reported).

Outline sketch

Prompt

Audience: RAG / search eng.
Steps:
1) Retrieve top-k candidates.
2) Per passage: Nouls relevant, usable_evidence, contradicts_premise, injection.
3) Code route(): accept / conflict / drop.
4) Build LLM prompt with separate accepted vs conflict blocks.
Expected effect: cleaner context than similarity rank alone (cookbook-reported).
Cross-link: Hands-on [RAG filtering](/guides/rag-filtering/).

Needs access to: typesafe-sdkyour-retriever

Who it's for

RAG / search engineers

Steps / how it's set up

Catalog card; Hands-on keeps the recipe. Sources: classifying_rag_passages · rerank_typesafe · awesome-jev-usecases · Search.

Suggested route order: injection → conflict → evidence sufficiency. Cross-link Hands-on RAG filtering.

Sources (wave-2 deepen)

Expected effect

Embeddings measure nearness; Jev asks whether a passage is usable evidence.

Unofficial outline for learning. Paraphrased from public docs and tutorials — not a production recipe. Review sources before you automate anything.