| name | rag-architect |
| description | Design, evaluate, and evolve production retrieval-augmented generation systems. Use when a task requires evidence-grounded search over private or specialized corpora, retrieval architecture selection, ingestion and permission design, hybrid search, reranking, citation integrity, offline evaluation, online monitoring, or diagnosis of retrieval failures. |
| zh_description | 用于以评测为先设计和诊断生产级 RAG，包括权限、混合检索、重排、引用和监控。 |
| version | 2.0.0 |
| author | seaworld008 |
| source | in-house |
| source_url | |
| tags | ["architect", "rag", "retrieval", "evaluation", "search"] |
| created_at | 2026-03-04 |
| updated_at | 2026-07-27 |
| quality | 5 |
| complexity | advanced |
RAG Architect
Design from measured retrieval failures, not from a fashionable vector
database or chunk size.
Start With the Product Contract
Define:
- users and authorization boundary
- answerable and unanswerable question classes
- source corpus, languages, formats, and update rate
- freshness and deletion requirements
- latency, cost, availability, and data-residency constraints
- citation and audit requirements
- acceptable abstention and false-answer rates
Do not build RAG when direct context, structured queries, a conventional search
index, or a deterministic API is simpler and more reliable.
Build the Evaluation Set First
Create a versioned dataset before tuning:
{
"id": "policy-017",
"query": "Can a contractor export customer data?",
"expected_source_ids": ["policy-access-4.2"],
"required_facts": ["contractors cannot export", "exception requires DPO approval"],
"forbidden_claims": ["contractors always have export access"],
"user_acl": ["policy:employee"],
"as_of": "2026-07-01",
