Retrieval accuracy, permission boundaries, and evaluation pipelines matter more than embedding model choice.
RAG demos work on small, curated document sets. Production RAG fails when knowledge is scattered, stale, permission-blind, or full of conflicting versions.
Hybrid retrieval
Hybrid retrieval with reranking consistently outperforms pure vector search on domain-specific enterprise queries. Golden datasets — updated when policies change — catch regression before users do.
Permission-aware access
Permission-aware access is non-negotiable. Agents must not retrieve documents the requesting user cannot access. This is an architecture problem, not a prompt trick.
What to measure
Measure retrieval accuracy, answer faithfulness, and business outcomes — not just latency and token cost.