After a public-health corpus migration, RAG answers pull off-topic SOPs even though the generator settings are unchanged. Which troubleshooting path best targets the likely cause?

Diagnosing embedding quality and drift, then re-vectorizing (or repairing) the migrated corpus so retrieval matches the new embeddings is the troubleshooting path that best targets off-topic SOP retrieval after a public-health corpus migration when generator settings are unchanged. Migrations often change embedding models, preprocessing, or vector spaces while leaving the generator alone; mismatched or drifted embeddings retrieve wrong SOPs even though generation parameters look identical. Repairing embeddings and re-indexing realigns retrieval with the migrated documents. Only raising temperature so off-topic SOPs sound more creative fails by concept: higher temperature increases randomness and does not fix retrieval of wrong SOPs. Disabling retrieval entirely and inventing public-health guidance without sources fails by concept and safety—ungrounded invention is worse than repairing RAG. Fixing the issue solely by changing the chat widget’s border radius fails by concept: UI cosmetics do not alter embedding space. Exam caveat: post-migration off-topic retrieval with unchanged generators implicates embedding/index health, not temperature knobs, retrieval abandonment, or UI tweaks. Operational check: compare embedding model IDs and dimensions pre/post migration, sample nearest neighbors for known public-health queries, re-embed and re-index if drift or mismatch appears, validate SOP retrieval on a golden set, then reopen generator tuning only if retrieval is healthy.