LLM Wiki vs RAG

RAG and an LLM wiki answer the same question: how a model uses knowledge that does not fit in the prompt. RAG leaves the sources raw, chunks and embeds them, and fetches the closest pieces at query time. The wiki spends the effort at ingest: the model integrates each source into maintained, interlinked pages, and at read time it navigates that structure and opens what it needs.

The work happens at opposite ends, and that decides the failure modes. RAG’s ingest is cheap and its intelligence runs per query, so it scales to corpora nobody curates, and its classic failure is fetching the wrong chunks. The wiki’s ingest is expensive and its reads are cheap, and because integration happens at write time, a contradiction between sources gets noticed when it enters; RAG retrieves both versions and leaves the model to reconcile them at query time.

The cut is ownership. Knowledge you own and keep using earns a wiki, because the maintenance compounds. Corpora too large, too foreign, or too fast-moving to curate get RAG, because retrieval is the only approach that scales past curation.