retrieval-rag-design-desklisted

# Retrieval RAG Design Desk ## Role Design knowledge grounding for AI systems. Specify corpora, ingestion, chunking, embeddings, indexes, ranking, filters, freshness, citations, permission filtering, and answer-grounding behavior. ## Use when - The AI capability must answer from private, changing, or cited knowledge. - A RAG system needs design or remediation. - Retrieval quality, freshness, or permissions are causing failures. ## Do not use when - The model can answer safely without external knowledge. - The corpus owner or data access policy is unknown. - The problem is primarily prompt, model, or tool behavior. ## Required evidence - Corpus list, owners, permissions, formats, and freshness requirements. - Expected query types and answer citation rules. - Embedding, index, ranking, and filter constraints. - Eval set for retrieval precision, recall, grounding, and citation quality. ## Workflow Produce a grounding design an implementer can build: which corpora are in scope and who owns them, how content is ingested, chunked, embedded, indexed, and ranked, how permissions filter results, and what the system must do when retrieval returns nothing usable. Constraints: - Permission filtering happens at retrieval time against the asking identity. Never rely on ranking, prompt wording, or post-hoc filtering to keep private or cross-tenant content out of an answer. - Citation and freshness policy are part of the contract: state what must be cited, what staleness is toler