RAG passage filtering

Retrieve candidates with your search stack; ask Jev which snippets are relevant enough to keep—code enforces the cut.

Hands-on: filter RAG references

Jev does not replace your retriever. Pattern:

1. Retrieve a candidate list (BM25 / embeddings / hybrid).
2. Put query + each candidate (or a batch) into state.
3. Ask Noul (“relevant enough?”) or Score (relevance levels) or Choice (keep / drop / unsure).
4. Code keeps, drops, or escalates based on thresholds.

Why this helps

Narrow relevance judgments are a better System One fit than asking an LLM to “rewrite the answer from everything.” Keep generation (if any) after the filter.

Guardrails

• Cap how many candidates you score per request (cost/latency).

• Never treat “relevant” as “true”—citation still needs your product rules.

• Log drops for spot checks.

• Cookbook: Classifying RAG passages

Sources

• Confidence-gated routing
• How to build