Excellent point. I completely agreeauthorization, tenant isolation, and document lifecycle should never depend on model-derived filters. Those are hard server-side constraints and should always fail closed.
I also like the idea of logging candidate counts after each hard filter. It makes debugging empty retrieval results much more transparent instead of leaving everyone guessing.
Thanks for the valuable addition!
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
The hard-versus-soft metadata distinction is especially important, but I would keep tenant, permission, and superseded-document gates outside any model-derived query filter. Soft constraints such as topic or freshness can degrade gracefully when extraction is uncertain; authorization and document status must fail closed under server-owned policy. Logging candidate counts before and after each hard gate also makes an empty retrieval result explainable instead of mysterious.