Step 7 — Bind AI providers and RAG to permission and quality

An AI feature is not one model call. Provider, model ID, embedding dimension, source documents, search copies, permissions, cost, and quality need separate contracts.

source → parse → embed → store → processed complete → search allowed → evidence recorded

An unprocessed file or another owner's vector must not become a fast but invalid result. DMDDKSL vector search filters on completed files, and HIRA joins and version selection preserve the same boundary.

An adapter translates provider failures into our retry and degraded-mode policy. Model changes need dimension checks, a small evaluation set, shadow comparison, and measurements for cost, latency, and missing evidence. A fallback that returns an empty success is misleading; preserve failure and retryability.

Define allowed evidence, freshness, and “no evidence” before optimizing answer prose. Operators need failed-file counts, model version, and latency.

Read next

• Expose completed files only in RAG
• Next step