Dmitry
1.0 KiB
1.0 KiB
LLM-Wiki
Summary
Instead of querying raw documents with RAG every time (and re-synthesizing from scratch), maintain a persistent, evolving wiki layer compiled from sources. The wiki becomes the primary working knowledge artifact: structured pages, links, flagged contradictions, and an accumulated synthesis.
Key idea
- RAG-only: retrieve chunks at query time → re-derive synthesis repeatedly.
- Compiled wiki: ingest sources once → extract + integrate → keep pages current.
The benefit is compounding: each new source updates the existing structure, so future questions are answered by traversing already-built pages and links.
What the LLM “owns”
- creating and updating wiki pages
- cross-linking entities/concepts
- flagging contradictions and superseded claims
- maintaining an index + a log
Humans focus on: sourcing, asking questions, deciding what matters.
Files
- Entry point: index
- Model: Architecture
- Process: Operations
- Seed source: 2026-04-12__llm-wiki__idea