The LLM Wiki: A Pattern for Smarter Organisational Knowledge Bases
Andrej Karpathy's 'LLM Wiki' pattern offers a compelling alternative to retrieval-augmented generation. Instead of re-deriving insights from scratch on every query, AI incrementally builds and maintains a persistent, structured knowledge base. The implications for organisational knowledge management are significant.
A Better Pattern for Organisational Knowledge
Andrej Karpathy - the AI researcher known for his work at OpenAI and Tesla - recently shared a pattern he calls the “LLM Wiki.” It addresses a fundamental limitation in how most organisations use large language models for knowledge work today.
The standard approach is retrieval-augmented generation (RAG): upload documents, the AI retrieves relevant chunks at query time, and generates an answer. It powers tools like ChatGPT file uploads and most enterprise search products. It works - but nothing accumulates. Every question forces the AI to rediscover knowledge from scratch, piecing together fragments across documents each time.
Ask a nuanced question that requires synthesising five sources, and the system rebuilds that synthesis on every single query. Nothing is ever built up.
There is a better pattern emerging, and it changes the economics of organisational knowledge management entirely.
The Core Idea
Instead of retrieving from raw documents at query time, you have the AI incrementally build and maintain a persistent wiki - a structured, interlinked collection of pages that sits between your team and the raw source material.
When a new source is added, the AI does not just index it for later retrieval. It reads the document, extracts key information, and integrates it into the existing knowledge base: updating entity pages, revising topic summaries, noting where new data contradicts old claims, and strengthening the evolving synthesis.
The knowledge is compiled once and kept current, not re-derived on every query. Cross-references are already in place. Contradictions have already been flagged. The synthesis already reflects everything that has been ingested. And the wiki gets richer with every source added and every question asked.
A Three-Layer Architecture
The pattern has a clean structure:
Raw sources are your curated collection of documents - articles, reports, transcripts, data files. These are immutable. The AI reads from them but never modifies them. This is your source of truth.
The wiki is a directory of AI-generated pages: summaries, entity profiles, concept explanations, comparisons, and an evolving overview. The AI owns this layer. It creates pages, updates them when new sources arrive, maintains cross-references, and keeps everything consistent.
The schema is a configuration document that tells the AI how the wiki is structured, what conventions to follow, and what workflows to use when ingesting sources or answering questions. You and the AI co-evolve this over time as you learn what works for your domain.
Three Operations That Make It Work
Ingest. When a new source is added, the AI reads it, writes a summary page, updates the index, and revises every relevant entity and concept page across the wiki. A single source might touch ten to fifteen pages. The knowledge compounds immediately.
Query. Questions are answered against the wiki rather than raw documents. Because the synthesis already exists, answers are faster, more comprehensive, and more consistent. Valuable answers - comparisons, analyses, connections discovered - get filed back into the wiki as new pages. Your explorations compound just like your sources do.
Lint. Periodically, the AI health-checks the wiki. It looks for contradictions between pages, stale claims superseded by newer sources, orphan pages with no connections, and important concepts that lack their own page. It can also suggest new questions to investigate and new sources to look for.
Why This Matters for Organisations
The tedious part of maintaining a knowledge base has never been the reading or the thinking - it is the bookkeeping. Updating cross-references, keeping summaries current, noting when new information contradicts old claims, maintaining consistency across dozens of pages. Teams abandon wikis because the maintenance burden grows faster than the value they provide.
This is a pattern we recognise from our own work. Organisations do not lack information. They lack the capacity to keep that information structured, current, and usable. The barrier is not knowledge creation; it is knowledge maintenance.
AI does not get bored, does not forget to update a cross-reference, and can touch fifteen files in a single pass. The wiki stays maintained because the cost of maintenance drops to near zero.
The human role shifts to what humans do best: curating sources, directing analysis, asking good questions, and deciding what it all means. This is precisely the division of labour we advocate for in any AI system - AI handling the mechanical work, humans providing the judgement and direction that determine whether the output is actually valuable.
Where This Connects to Tacit Knowledge
There is an important extension of this pattern that Karpathy does not explicitly address but that follows naturally from the architecture: when the sources being ingested are not just documents, but records of expert reasoning - transcripts of experienced practitioners explaining their decisions, notes from operational debriefs, captured reflections on why a particular approach worked or failed - the wiki becomes something more than a knowledge base. It becomes a structured, searchable, evolving repository of expertise that would otherwise exist only in the minds of the people who developed it.
This is where the pattern intersects directly with knowledge transfer. The same mechanisms that make the LLM Wiki effective for synthesising published research make it equally effective for preserving and organising the tacit knowledge that experienced practitioners carry. The AI handles the structuring, cross-referencing, and maintenance. The experts provide the raw material - their reasoning, their judgement, their hard-won understanding of how things actually work.
A Practical Caveat
The AI is the writer, but a human should remain the editor-in-chief. In high-stakes contexts especially, build source citation into your process and budget time to verify the wiki’s claims against the original sources. The pattern works best when it amplifies human judgement rather than replacing it - a principle that holds for any AI system, but is particularly important when the knowledge being managed has real operational consequences.
Where This Applies
The pattern is domain-agnostic. It applies to competitive analysis and market intelligence, due diligence processes, research synthesis across weeks or months of reading, internal team knowledge bases fed by meeting transcripts and project documents, and ongoing learning initiatives where knowledge needs to be organised rather than scattered.
The key question for any knowledge-intensive process is straightforward: are you re-deriving insights from scratch every time, or are you building on what you have already learned? If the answer is the former, this pattern is worth exploring.
Karpathy’s full write-up is available as a public GitHub gist. The pattern is open and implementation-agnostic - the right version for your organisation depends on your domain, your sources, and the expertise you are trying to preserve.
If you are thinking about how to structure AI around your organisation’s knowledge and expertise, our AI strategy consulting is designed for exactly that conversation. Get in touch.
FAQ
Frequently asked questions
What is the LLM Wiki pattern and how does it differ from RAG?
The LLM Wiki is a pattern proposed by AI researcher Andrej Karpathy in which an AI incrementally builds and maintains a persistent, structured wiki from raw source material, rather than retrieving and re-synthesising document fragments on every query. In standard retrieval-augmented generation (RAG), every question forces the system to rediscover knowledge from scratch. With the LLM Wiki, knowledge is compiled once, kept current, and enriched over time, so that cross-references, contradictions, and synthesis are already in place when a question is asked.
How does the LLM Wiki pattern relate to organisational knowledge management?
The pattern addresses a core challenge in knowledge management: the gap between having information and having usable, structured knowledge. Organisations accumulate vast quantities of documents, but the work of synthesising, cross-referencing, and maintaining that knowledge is where most efforts fail. The LLM Wiki shifts the maintenance burden to AI, allowing human effort to focus on curation, direction, and judgement, the parts of knowledge management where people add the most value.
What role do humans play in an LLM Wiki system?
Humans serve as editor-in-chief. They curate which sources are ingested, direct the analysis by asking the right questions, define the schema that governs how the wiki is structured, and verify the AI's claims against original sources. The AI handles the bookkeeping: updating cross-references, flagging contradictions, maintaining consistency. This division of labour plays to the strengths of each, and ensures that expert judgement remains central to how organisational knowledge is shaped and used.
Work with us
Want to apply this in your organisation?
Tell us what you are working on and we will map practical next steps for your team.