Definition

An LLM knowledge base is a workflow where immutable source material is incrementally compiled by an LLM into a persistent, interlinked markdown wiki.

Why It Is Different From Plain RAG

  • Plain RAG retrieves raw fragments at question time.
  • This method maintains a durable synthesis ahead of time.
  • Cross-links, summaries, and contradictions accumulate instead of being rediscovered for every query.

Working Model In This Repo

  • Source material lives outside wiki/ and is treated as immutable.
  • wiki/sources/ holds source summaries.
  • wiki/concepts/ holds cross-source synthesis.
  • wiki/queries/ and wiki/reports/ hold durable outputs from asking questions.
  • wiki/index.md and wiki/log.md help navigation and maintenance.

Contrast With AI Knowledge Bases

  • Many AI knowledge base products focus on question-time grounding over uploaded documents.
  • An LLM-maintained wiki adds a persistent synthesis layer: summaries, cross-links, contradiction tracking, and updated concept pages.
  • The two approaches can complement each other: an AI knowledge base can help with retrieval, while the wiki preserves durable synthesis.

Current Evidence In Repo

Practical Loop

  1. Add a source to one of the immutable source folders outside wiki/.
  2. Ingest it into wiki/sources/.
  3. Update any affected concept pages.
  4. Ask questions against the wiki.
  5. Save useful answers back into the wiki.
  6. Periodically lint for missing links, stale claims, and gaps.
  • Pinecone Nexus 将”预编译知识”产品化为企业基础设施：Context Compiler 自动构建 artifact，KnowQL 提供声明式查询，KRAFTBench 量化了编译式 vs 查询时组装的差距（token 成本差 7-80 倍）。本仓库的 LLM-maintained wiki 可视为同一思路的小型个人版。

Early Takeaway

Even at small scale, this method can be useful because the wiki index and concise summaries reduce repeated search effort and make later questions easier to answer consistently.

Evidence 区补充 Karpathy 演讲摘要、llm-training-truth、ynarwal-how-llms-work-visual 与 large-language-model-fundamentals 互链；2026-05-10。