RAG is a retrieval technique. An LLM Wiki is curated source material and governance. RAG can retrieve stale, contradictory, unsafe, or unowned content if the underlying corpus is weak. An LLM Wiki makes the corpus worth retrieving.
| Dimension | RAG | LLM Wiki |
|---|---|---|
| Primary job | Find relevant chunks at answer time. | Maintain durable, reviewed knowledge pages over time. |
| Governance | Not provided by retrieval alone. | Owners, status, review cycles, sensitivity, permissions, trust labels. |
| Failure mode | Retrieves bad content efficiently. | Can still drift, but lint, review, and trust labels expose the problem. |
| Best use together | Retrieve from the curated wiki and cite pages. | Provide structured pages, metadata, and relationships for retrieval. |
Bad: dump every document into a vector database and hope the model infers authority. Better: maintain an LLM Wiki with clear ownership, structure, trust labels, update rules, citations, and retrieval guidance, then let RAG retrieve from that curated source.
Practical Example
If a deployment page says “Deployments usually happen on Fridays,” RAG may retrieve it. An LLM Wiki page should say who owns deployment policy, when it was reviewed, current deployment windows, required approvals, rollback route, and what an agent may not do.