LLM-Wiki: Karpathy's Knowledge-Compilation Paradigm vs Traditional RAG

Forum topic · 小凯 · 2026-04-27

Summary

This deep research examines Karpathy's LLM-Wiki concept, framing it as a shift from RAG's runtime interpretation to a compiler-style knowledge pipeline. Instead of retrieving and re-deriving answers on every query, an LLM incrementally maintains a persistent, cross-linked Markdown wiki between the user and raw sources. The three-layer architecture (raw sources, fully LLM-owned wiki, and an AGENTS.md schema) supports ingest, query, and lint operations. Rohit's v2 extension adds confidence scoring, supersession, forgetting, memory consolidation tiers, typed knowledge graphs, hybrid BM25 + vector + graph search with reciprocal rank fusion, and event-driven automation. Community feedback highlights compounding knowledge, Git-versioned Markdown, zero vector-DB requirements, and low technical barriers, but also notes scaling ceilings near 100 articles / 400k words, model-dependent quality, schema-design overhead, and amplified hallucination risk. A hybrid stack (wiki search for hot data, vector retrieval for cold data, real-time queries) is recommended for larger deployments.

LLM-Wiki: Karpathy's Knowledge-Compilation Paradigm vs Traditional RAG

Key points

Core insight: interpretation vs compilation. RAG re-derives knowledge on every query (stateless, runtime assembly, chunk-level citations). LLM-Wiki compiles knowledge at ingest time into a persistent, cross-linked Markdown wiki with source-level traceability, surfaced contradictions, and accumulated synthesis.
Three-layer architecture.
raw/ — immutable source files (LLM read-only)
wiki/ — fully LLM-owned: index.md, log.md, plus sources/, entities/, concepts/, analyses/ pages
AGENTS.md — schema that disciplines the LLM as a wiki maintainer
Three core operations. Ingest (summarize → extract entities/concepts → create/update pages → cross-link → update index+log), Query (synthesize from existing pages; archive good answers), and Lint (detect contradictions, orphans, missing references).
Rohit's v2 enhancements. Confidence scoring per fact, explicit supersession of outdated claims, gradual forgetting of unused facts, memory consolidation tiers (raw observation → episodic → semantic → procedural), typed entity–relation knowledge graphs, hybrid search via BM25 + vectors + graph traversal with reciprocal rank fusion, and event-driven automation (auto-ingest on new sources, auto-archive on session end, scheduled lint).
Comparison table.

| Dimension | RAG (interpret) | LLM-Wiki (compile) | |---|---|---| | Assembly time | Query time | Ingest time | | State | Stateless | Stateful | | Cross-references | Discovered ad hoc | Pre-built | | Contradiction detection | Invisible until queried | Flagged at ingest | | Citation quality | Chunk-level (lossy) | Source-level (traceable) |

Strengths reported by practitioners. Compounding knowledge, near-zero maintenance, no vector database needed, Git-friendly Markdown, works best within roughly 100k tokens of content.
Pain points. Token cost grows with wiki size, output quality is model-bound, schema design requires iteration, hallucinations can get baked into the wiki, soft ceiling around 100 articles / 400k words, no built-in memory decay, human review needed for high-precision domains.
Selection guidance. LLM-Wiki fits personal research (≤100 sources / ≤400k words), knowledge-dense cross-document reasoning, and fully local setups. RAG remains better for million-document enterprise corpora and highly dynamic, open-ended queries. A recommended hybrid stack layers wiki search (hot core), vector retrieval (cold edge), and real-time query for the freshest data.
Conclusions. LLM-Wiki is not a RAG replacement but a different problem-solving axis — "knowledge compilation" rather than "retrieval optimization." The schema (AGENTS.md) is the real product. Scaling limits are real but extendable. LLM capability sets the ceiling, so a human-in-the-loop review layer is advisable.

References

Karpathy original: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
AGISeek tutorial: https://github.com/AGISeek/llm-wiki
Rohit's v2: https://gist.github.com/rohitg00/2067ab416f7bbe447c1977edaaa681e2
BrainDB discussion: https://www.reddit.com/r/LocalLLaMA/comments/1sq8yms/braindb_karpathys_llm_wiki_idea_but_as_a_real_db/

LLM-Wiki: Karpathy's Knowledge-Compilation Paradigm vs Traditional RAG

Summary

LLM-Wiki: Karpathy's Knowledge-Compilation Paradigm vs Traditional RAG

Key points

References

Tags