You believe that the retrieval harm curve exists in you...

Explicit relevance labels are not required to locate the crossover point where retrieval starts harming performance because final answer correctness serves as the label. To find this threshold:

1. Sample a few thousand historical production queries.
2. Execute both branches offline: closed-book (parametric only) and RAG (retrieval-augmented).
3. Evaluate generated outputs against available downstream outcome signals (e.g., resolved customer support tickets, user feedback/thumbs, or an LLM judge calibrated against a held-out set of human labels).
4. Bucket queries into bins based on entity popularity.
5. Identify the popularity bucket where closed-book accuracy matches or exceeds RAG accuracy.

If standard popularity metrics (like Wikipedia pageviews) are unavailable for domain-specific data, use the frequency of the query’s head noun phrase within your proprietary internal corpus as a proxy for parametric familiarity.