Skip to main content

​
Compounding Knowledge Base

A knowledge base where each new source and each query adds to a persistent, interlinked structure — rather than being processed in isolation at retrieval time.

​
The core distinction: compilation vs. retrieval

RAG (Retrieval-Augmented Generation): raw documents sit indexed; at query time the LLM retrieves relevant chunks and generates an answer from scratch. Nothing is built up. Ask a question requiring synthesis across five documents and the LLM must find and piece together fragments every time. Compounding knowledge base: the LLM compiles knowledge incrementally as sources arrive. Cross-references are already built. Contradictions already flagged. Synthesis already reflects everything ingested. Each new source strengthens or revises the existing structure rather than adding to a pile. The difference is analogous to an interpreted vs. compiled program: RAG interprets source documents on every query; the compounding wiki compiles them once and keeps the result current.

​
Why it compounds

  • Ingested sources → summary + entity/concept page updates (one source touches many pages)
  • Answered queries → good answers get filed back as new pages; explorations accumulate
  • Lint passes → contradictions resolved, gaps identified, cross-references added
The wiki accretes value from all three operations, not just from ingestion.

​
Relationship to compound engineering

Compound Engineering applies the same compounding principle to software delivery. A compounding knowledge base turns sources and questions into durable understanding; compound engineering turns tasks and reviews into durable capability. The shared rule is: do not let useful work evaporate after one session.

​
Why humans abandon wikis

The bottleneck isn’t reading or thinking — it’s bookkeeping. Updating cross-references, keeping summaries current, noting when new data contradicts old claims, maintaining consistency across dozens of pages. This cost grows faster than value. LLMs remove the bottleneck: they don’t get bored, don’t forget a cross-reference, can touch 15 files in one pass.

​
The human’s role

Curation, direction, and synthesis judgment. Specifically: which sources to ingest, what questions to ask, whether the LLM’s synthesis is right, where to push deeper. The LLM handles all the maintenance.

​
RAG nuance

Note (2026-04-22): Contextual Retrieval significantly improves RAG — −49% retrieval failure with contextual embeddings + BM25, −67% with reranking. The contrast above (“nothing accumulates”) remains valid for standard RAG but overstates the ceiling. Improved RAG is a legitimate option for knowledge bases that don’t need the cross-source synthesis a compiled wiki provides. The distinction is still real: RAG answers queries from raw sources; a compounding wiki builds durable structure. But the gap is smaller than implied when RAG is properly implemented.
  • Llm Wiki Pattern — the specific implementation of this pattern described by Karpathy
  • qmd — search tool that enables efficient query operations as the wiki scales
  • Contextual Retrieval — technique that significantly narrows the RAG/wiki performance gap for retrieval tasks
  • Compound Engineering — workflow-level version of the same compounding principle