Teams ask us to swap models when their assistant answers badly. We ask to see the corpus instead. Nine times out of ten we find duplicated policies with conflicting dates, tables flattened into mush, and documents whose only retrieval-relevant sentence lives in a footer.
Retrieval quality is decided long before the embedding step: deduplication, chunking that respects document structure, metadata that captures authority and recency. A mediocre model over a clean, governed corpus beats a frontier model over a landfill.
Our rule: no RAG project starts with a model choice. It starts with a corpus audit and a definition of which document wins when two disagree. That decision is organizational, not technical — and it's the one the software can't make for you.