Breaking Business News | Breaking business news AM | Breaking Business News PM | Business News Select | Link from bio | SMPostStory | Terms

“An LLM Wiki incrementally builds and maintains a persistent wiki – a structured, interlinked collection of markdown files that sits between you and the raw sources instead of just retrieving from raw documents at query time…. the wiki is a persistent, compounding artifact.” – LLM Wiki – Andrej Karpathy

An LLM Wiki is a system where a large language model (LLM) incrementally builds and maintains a persistent wiki-a structured, interlinked collection of markdown files-that serves as an intermediary knowledge layer between raw source documents and user queries, rather than relying solely on real-time retrieval from unstructured data.¹ This approach transforms ad-hoc document processing into a compounding artifact, where new information integrates into an evolving graph of summaries, concept pages, entity profiles, comparisons, and syntheses, enabling deeper reasoning over accumulated knowledge.^1,3,6

Core Mechanics of Compilation and Persistence

The process begins with raw sources-articles, papers, repositories, datasets, or images-dropped into a designated directory.^3,6 The LLM acts as a compiler, processing only modified files incrementally to update the wiki without full recompilation.^1,6 It generates markdown files including an index with summaries, dedicated pages for concepts and entities, cross-references via wiki-links, and derived outputs like charts or slides.^1,3,12 This persistence addresses LLM limitations such as hallucinations and outdated knowledge by creating a grounded, human-readable knowledge substrate that grows with each addition.^2,11

• Raw ingestion: Unstructured inputs land in a raw/ folder, triggering selective LLM processing.³
• Wiki generation: LLM produces wiki/ structure with INDEX.md for navigation, concepts/ subfolders for topical articles (~100 pages at scale, totaling around 400 000 words), and backlinks.^3,6
• Query interface: Users query the wiki via agents; responses generate new markdown, slides (e.g., Marp format), or visuals (e.g., matplotlib), filed back to enhance the base.^6,9
• Tools integration: Obsidian serves as the frontend for browsing; naive search or CLI tools handle retrieval without vector databases.^6,15

At moderate scale, this eliminates the need for embeddings or complex retrieval pipelines, as the LLM reads the index (a few thousand tokens) and relevant pages directly.^6,12

Practical Implications in Knowledge Workflows

In practice, an LLM Wiki shifts token usage from code manipulation to knowledge manipulation, supporting research on topics like AI scaling laws or specific domains.³ For a researcher, it means querying multi-step questions-e.g., comparing 50 papers on a topic-that would take hours manually, now answered in context of ~400 000 words of synthesized content.^3,6 Outputs persist, compounding value: a query on contradictions flags inconsistencies during ingestion, not ad-hoc at query time.¹²

Directory structure exemplifies simplicity:

my-research/
raw/          # Sources
wiki/          # LLM-owned
  INDEX.md
  concepts/
    concept-a.md
output/        # Query artifacts
_meta/         # State

This yields a durable, editable artifact versus ephemeral chat responses, with the LLM owning maintenance for consistency.^1,12

Contrast with Retrieval-Augmented Generation (RAG)

Traditional RAG retrieves relevant documents at query time via vector search, augmenting prompts to mitigate LLM gaps in domain knowledge, factuality, or recency.^2,5,11 While effective for dynamic, knowledge-intensive tasks, RAG processes knowledge per query, lacking persistence or pre-built connections.¹²

¹²

RAG excels in scale (e.g., enterprise databases) with techniques like hybrid search, recursive retrieval, or post-retrieval reranking.² LLM Wiki suits personal or moderate-scale bases (~100 articles), prioritizing structure over speed.^6,12 Advanced RAG variants (e.g., RETRO, Self-RAG) introduce iteration or adaptation, converging toward wiki-like persistence.²

Major Schools of Thought in LLM Knowledge Management

Two paradigms dominate: dynamic retrieval (RAG lineage) and static compilation (wiki-style).^11,12 RAG, originating from 2020 research, emphasizes external augmentation without retraining, popular in chatbots and domain apps.^5,14 Compilation approaches treat LLMs as builders of structured artifacts, echoing knowledge graphs or personal wikis like Roam/Obsidian, but automated.⁶

• Dynamic retrieval school: Prioritizes real-time access; variants include naive RAG (basic fetch-generate), advanced (query expansion, reranking), and modular (self-improvement).^2,11
• Persistent compilation school: Builds durable structures upfront; LLM Wiki exemplifies, with incremental updates and link graphs.^1,6
• Hybrid evolution: Emerging methods blend, e.g., RAG with memory or iterative retrieval.^2,5

Leading Theorists and Proponents

Andrej Karpathy, former OpenAI/Tesla AI director, formalized LLM Wiki in April 2026 via GitHub Gist, describing it as his primary workflow for research knowledge bases.^1,3 His insight: at personal scale, structured markdown suffices over vector search, with LLMs handling interlinking.^1,12 Karpathy’s ~100-article wikis on topics demonstrate scale, using tools like Obsidian and agentic Q&A.^3,6,15

Broader RAG theorists include Patrick Lewis (RETRO co-author) and teams at Google/DeepMind, advancing retrieval paradigms.^2,5 Implementers like Databricks and Google Cloud promote RAG for enterprise.^8,14 Community breakdowns (e.g., antigravity.codes, DAIR.AI) dissect Karpathy’s system, providing diagrams and minimum viable setups.^3,6

Tensions and Debates

Scale limits one tension: LLM Wiki thrives at 400 000 words but may falter beyond without indexing aids; RAG scales via vectors.^6,12 Transparency versus efficiency pits editable markdown against black-box retrieval.^1,5 Incremental compilation risks drift if LLM generations inconsistent, though human oversight (reading, not editing) mitigates.^3,9

Debate swirls on obsolescence: as LLMs grow (e.g., models with 10^24 FLOPs training¹³), need for external bases diminishes, yet domain/recency gaps persist.^10,11 Cost: wiki compilation consumes upfront tokens (e.g., processing 100 docs), RAG defers to queries.² Evaluation lacks standards; RAG benchmarks exist, but wiki efficacy is anecdotal.¹¹

• Hallucination: Wiki flags via structure; RAG via grounding.^2,12
• Update latency: Wiki incremental; RAG instant.⁵
• Editability: Wiki human-readable; RAG opaque.¹

Strategic Relevance Today

LLM Wiki matters as knowledge work surges-researchers, analysts face info overload amid 2T-token training corpora.^10,13 It enables compounding intelligence: each query enriches the base, yielding multi-hop reasoning impossible in raw RAG.^3,6 In 2026, with models like Phi-2 (2,7B params, 1,4T tokens¹³), personal bases bridge proprietary gaps.

For teams, it prototypes team wikis; enterprises adapt for compliance via auditable markdown.¹² As LLMs evolve, wiki patterns influence agentic workflows, where compilation precedes action.⁶ The persistent artifact endures, outlasting query sessions, positioning it as a foundational tool in AI-augmented cognition.

Implementation Considerations

Start minimal: script LLM prompts for index/concept generation, use git for versioning.³ Scale with agents for auto-filing.⁶ Challenges include prompt engineering for consistency (e.g., “maintain backlinks”) and storage (400 000 words ~few MB).¹ Future: integrate with LLMs supporting tools for native compilation.

This system redefines RAG by front-loading structure, offering a practical path to persistent, queryable knowledge in an era of exploding data.

References

1. LLM WIki – https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

2. LLM – Extensions Wiki (XWiki.org) – 2026-03-18 – https://extensions.xwiki.org/xwiki/bin/view/Extension/LLM/

3. Retrieval Augmented Generation (RAG) for LLMs – 2026-02-01 – https://www.promptingguide.ai/research/rag

4. Karpathy’s LLM Knowledge Bases: The Post-Code AI Workflow – 2026-04-03 – https://antigravity.codes/blog/karpathy-llm-knowledge-bases

5. Large language models – Wikiversity – 2025-12-24 – https://en.wikiversity.org/wiki/Large_language_models

6. Retrieval-augmented generation – Wikipedia – 2023-11-05 – https://en.wikipedia.org/wiki/Retrieval-augmented_generation

7. LLM Knowledge Bases – DAIR.AI Academy – 2026-04-03 – https://academy.dair.ai/blog/llm-knowledge-bases-karpathy

8. llm-wiki – GitHub Gist – 2026-04-04 – https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f?permalink_comment_id=6079056

9. What is Retrieval-Augmented Generation (RAG)? – Google Cloud – https://cloud.google.com/use-cases/retrieval-augmented-generation

10. LLM Knowledge Bases post by Andrej Karpathy – DeepakNess – 2026-04-03 – https://deepakness.com/raw/llm-knowledge-bases/

11. Large language model – Wikipedia – 2023-03-09 – https://en.wikipedia.org/wiki/Large_language_model

12. Retrieval-Augmented Generation for Large Language Models – arXiv – 2023-12-18 – https://arxiv.org/abs/2312.10997

13. Karpathy’s LLM Wiki: The Complete Guide to His Idea File – 2026-04-04 – https://antigravity.codes/blog/karpathy-llm-wiki-idea-file

14. List of large language models – Wikipedia – 2023-03-09 – https://en.wikipedia.org/wiki/List_of_large_language_models

15. What is Retrieval Augmented Generation (RAG)? – Databricks – 2023-10-18 – https://www.databricks.com/blog/what-is-retrieval-augmented-generation

16. Andrej Karpathy’s LLM-powered personal knowledge base workflow … – 2026-04-03 – https://www.youtube.com/watch?v=VLd0K0bkOIE