It sounds like you've hit on a common challenge with semantic search systems, especially when moving from controlled environments to real users with vague queries. We've seen this exact scenario before where the architectural and product-level decisions around interpretation and retrieval become far more critical than the underlying models. This usually comes down to how the system handles ambiguity and inferred constraints, often leading to empty states or irrelevant results if not designed carefully. Happy to sanity-check your approach or share insights on how other teams have navigated similar issues with hybrid SQL + vector stacks.