RAG: Governed Retrieve, Validated Pack

This is the reference design for retrieval in production. Principles live in G.A.I.N RAG. Playbooks live under RAG. Policy verdicts stay in Governance Blueprint.

THE CLAIM

The route declares which corpora this capability may touch. The app fetches. The LLM synthesizes from a validated pack. Search is not permission.

What you are building

A governed retrieval path on one request:

1. Route pin: retrieval.mode and retrieval.scope from the route row
2. Who starts retrieve: prefetch (app), named stage (workflow), or tool (LLM proposes inside allowlist)
3. PEP: action is retrieve; resource is one corpus in scope
4. Gateway pack: ACL, rerank, token budget, source ids
5. Validate, then synthesize: the app initiates validation; the model never searches the index directly

Customer prefs and session summaries are not this path. They are the memory service: Memory.

Control points

Control	Lives on	Must not
retrieval.mode	Route row	Be inferred from the prompt
retrieval.scope	Route row	Widen mid-session
Corpus ACL	Retrieval gateway	Be a "don't search HR" instruction
PEP verdict	Governance	Be skipped for "read-only" search

How the app branches per mode and pattern: RAG retrieval.

Eval and ownership

Eval: Plane ③ Context. RAG / knowledge teams own corpora and packs. Security owns entitlements on the corpus resource. AI platform owns the pin and the PEP call.

Playbooks

• RAG overview · RAG retrieval
• Runtime for PEP/PDP
• Route contract for the retrieval field