LLM 维基

Problem

RAG answers questions but never accumulates understanding. Every query starts from scratch and nothing gets corrected over time.

Approach

A persistent, interlinked markdown wiki that a coding agent maintains, not a vector index. Three layers: raw sources, wiki pages, and a schema that tells the agent how pages must look.

• acquire fetches and normalizes sources (docs, notebooks, PDFs) with change detection, so unchanged sources are skipped.
• ingest has the agent write or update pages, keeping every claim linked back to its source.
• lint runs deterministic checks (broken links, missing sources, orphans) and an agent pass for contradictions.

The public edition targets Google’s open-source Meridian marketing-mix-modeling library as its corpus.

Result

A knowledge base that compounds: each ingest makes the next answer better, and every page cites where it came from.