Take into consideration revisiting objects you’ve saved to Pocket, Notion or your bookmarks. Most individuals don’t have the time to re-read all of this stuff after they’ve saved them to those varied apps, except they’ve a necessity. We’re wonderful at accumulating tons of data. Nevertheless, we’re simply not superb at making any of these locations work together with one another or add a cumulative layer that connects them collectively.
In April of 2026, Andrej Karpathy (former AI Director of Tesla and co-founder of OpenAI) prompt an answer to this situation: use a big language mannequin (LLM) to construct your wiki in real-time.
This concept turned viral and was finally adopted with a GitHub gist describing how to do that utilizing an LLM. This information will present the entire help (with instance code) for constructing your individual residing, evolving private wiki.
The RAG Downside: Rediscovering Data from Scratch, Each Single Time
Numerous trendy AI data instruments use Retrieval-Augmented Technology (RAG). In easy phrases, you add paperwork, ask a query, and the system retrieves related textual content to assist the AI generate a solution. Instruments like NotebookLM, ChatGPT (file uploads), and most enterprise AI techniques observe this strategy.
On the floor, this is sensible. However as Andrej Karpathy factors out, there’s a key flaw: the mannequin doesn’t accumulate data. Every question begins from scratch.
If a query requires synthesizing 5 paperwork, RAG pulls and combines them for that one response. Ask the identical query once more tomorrow, and it repeats your entire course of. It additionally struggles to attach data throughout time, like linking an article from March with one from October.
Briefly, RAG produces solutions, but it surely doesn’t construct lasting data.
Wiki Resolution: Compile Data As soon as, Question Eternally
Karpathy’s strategy will change the way in which we take a look at fashions. Reasonably than processing uncooked paperwork after our queries, we are going to now course of these paperwork on the time of ingestion. The outcomes of this processing will probably be a everlasting, structured wiki-like product (which is able to help you retailer and retrieve paperwork with a excessive diploma of management over their location).
While you add a brand new doc supply to the LLM, the LLM doesn’t merely create an index of that supply for later retrieval. As a substitute, the LLM reads, understands, and integrates that supply into the data base, updating all related present pages (the place mandatory). It notes down any contradictions between the brand new and present claims or data, creating any mandatory new idea pages, and reinforcing the complicated relationships throughout your entire wiki.
In response to Karpathy, “With LLMs, data is created and maintained repeatedly and persistently slightly than with the ability to use particular person queries to create data.” Right here is a straightforward comparability that illustrates this distinction additional.
Dimension
Conventional RAG
LLM Wiki
When data is processed
At question time (each single query)
At ingest time (as soon as per supply)
Cross-references
Found ad-hoc, or missed totally
Pre-built and maintained
Contradictions between sources
Usually not seen
Flagged throughout ingestion
Data accumulation
Zero — begins recent every question
Compounds with each supply added
Output format
Ephemeral chat responses
Persistent, editable Markdown information
Knowledge possession
Inside a supplier’s system
In your native laptop
How It Really Works: A Step-by-Step Information
Let’s evaluation how a person would develop certainly one of these wikis.
Step 1: Acquire your sources
It’s worthwhile to accumulate every little thing – articles that you’ve saved, books loved, notes you will have created, transcripts from discussions, and even your very personal historic conversations. All these supplies are your uncooked supplies, simply as ore should bear refining earlier than use.
Among the finest practices from this group is to not deal with all paperwork in the identical trend. For instance, a 50-page analysis white paper requires extraction on a section-by-section foundation whereas a tweet or social media thread solely requires a main perception and corresponding context. Likewise, a gathering transcript requires extraction of selections that had been made, motion objects which can be to be carried out and key quotations. By first classifying the kind of doc will assist extract the fitting kind of data to the correct quantity of element.
Step 3: The AI writes wiki pages (Question)
You’ll feed your supply supplies into your AI’s LLM through a structured question. It should permit the LLM to provide a number of wiki pages that conform to the established template of getting: a frontmatter block (YAML), a TLDR sentence, the physique of the content material, and the counterarguments/knowledge gaps.
Step 4: Create an index
A central index.md will function a desk of contents, and hyperlink straight to every web page of the wiki. That is how an AI agent can effectively traverse your entire data base; it begins on the index, reads by the tldr’s, then drills down into solely these pages which can be related to its particular query.
Step 5: Document your questions
This is without doubt one of the most under-appreciated options of the system. While you ask the LLM a well-formed query and obtain a response that gives precious perception. For instance, a comparability between two frameworks, or a proof of how two ideas are associated, you save that response as a brand new wiki web page tagged with the label query-result. As time goes on, your greatest pondering has been collected slightly than misplaced in chat logs.
Step 6: Conduct lint passes
At applicable intervals, you ask the LLM to audit your entire wiki for contradictions or inconsistencies between pages, and to point these statements which have been rendered out of date by a more moderen supply. Moreover, the LLM will present enter on figuring out orphan pages (i.e., pages that don’t have any hyperlinks pointing to them), and for offering an inventory of ideas which can be referenced inside the present content material however are usually not but represented by their very own respective pages.
Karpathy talks about numerous particular instruments in his Gist. Beneath you can find what every instrument does and the way they match into his total workflow.
1. Obsidian – Your Wiki IDE
Obsidian is a free markdown data administration utility which makes use of a neighborhood listing as a vault. For Karpathy, that is the viewing interface used for the wiki as a result of it has three distinct options that matter for his system:
- The Graph View offers a graph of all wiki pages represented as nodes, and as well as, each wiki hyperlink ( [[wiki-links]] ) will probably be represented as edges connecting all nodes collectively. Hub pages will probably be related to many nodes, and so will probably be represented as bigger than common nodes. Orphan pages will probably be represented as remoted nodes. This permits for fast visible illustration of the density of information and gaps inside an individual’s data. No different doc view or file browser can present this illustration visually.
- The Dataview Plugin permits customers to show their wiki right into a database that may be queried. All pages will need to have yaml frontmatter, so the combination specification is glad and subsequently permits the person to run SQL-like queries towards all pages within the wiki.
# In any Obsidian observe, Dataview renders this dynamically:
# Record all idea pages ordered by variety of sources
TABLE size(sources) AS “Supply Rely”, confidence, up to date
FROM “wiki/ideas”
SORT size(sources) DESC
# Discover low-confidence pages that want evaluation
TABLE title, sources
FROM “wiki”
WHERE confidence = “low”
SORT file.mtime ASC
# Discover pages not up to date within the final 2 weeks
LIST
FROM “wiki”
WHERE up to date < date(right this moment) – dur(14 days)
SORT up to date ASC
- The Net Clipper browser extension (obtainable for Chrome, Firefox, Safari, Edge, Arc, and Courageous) converts internet articles to wash Markdown with YAML frontmatter in a single click on, saving on to your uncooked folder. You obtain all photos to your laptop by urgent the hotkey Ctrl+Shift+D after you end clipping as a result of the LLM requires entry to the pictures.
2. Qmd: Search at Scale
The LLM can use the index.md file to entry the wiki content material with out issues at small scale. The index turns into unreadable in a single context window when you will have greater than 100 pages as a result of it reaches extreme measurement.
The native search engine qmd allows Markdown file searches by three search strategies which Tobi Lutke (CEO of Shopify) developed. The system operates totally in your machine as a result of it makes use of node-llama-cpp with GGUF fashions which require no API connections and shield your knowledge from leaving your laptop.
# Set up qmd globally
npm set up -g @tobilu/qmd
# Register your wiki as a searchable assortment
qmd assortment add ./wiki –name my-research
# Fundamental key phrase search (BM25)
qmd search “combination of specialists routing effectivity”
# Semantic search, finds associated ideas even with totally different phrases
qmd vsearch “how do sparse fashions deal with load balancing”
# Hybrid search with LLM re-ranking, highest high quality outcomes
qmd question “what are the tradeoffs between top-k and expert-choice routing”
# JSON output for piping into agent workflows
qmd question “scaling legal guidelines” –json | jq ‘.outcomes[].title’
# Expose qmd as an MCP server so Claude Code can use it as a local instrument
qmd mcp
The MCP server mode allows Claude Code to make use of qmd straight as a built-in instrument which ends up in smoother workflow integration all through your knowledge ingestion and question processing duties.
3. Git: Model Management for Data
As a result of your total wiki is a folder of plain Markdown information, you get model historical past branching and collaboration free of charge with Git. That is fairly highly effective:
# Initialize the repo once you begin
cd my-research && git init
# Commit after each ingest session
git add .
git commit -m “ingest: MoE effectivity article — flags dense-vs-sparse contradiction”
# See precisely what modified in any ingest
git diff HEAD~1
# Roll again a nasty compilation move
git revert abc1234
# See how your data advanced over time
git log –oneline wiki/ideas/mixture-of-experts.md
# Share with a staff through GitHub (the wiki turns into collaborative)
git distant add origin https://github.com/yourname/research-wiki
git push -u origin predominant
Getting Began: Your First LLM Wiki in Three Steps
In case you’re enthusiastic about this idea there’s a straightforward option to start:
- Choose one space of curiosity you might be at the moment exploring and provides the AI 5-10 of your greatest sources. Don’t try to put every little thing you’ve completed digitally into one place on the primary day however as a substitute learn the way the system works and how you can apply it to a small scale.
- Create the essential framework quickly. Create a wiki/listing on your wiki and have an index.md file there. Write down what your frontmatter is (title, kind, supply, created, up to date, tags), and be constant in naming your information e.g., concept-name.md or firstname-lastname.md. If this isn’t completed it will likely be troublesome to rectify later.
- Spend lots of time creating your preliminary immediate. That is essentially the most crucial step. Create guidelines for Classifying, writing TLDRs, writing the frontmatter in addition to guides for when to create Pages and when to edit the pages. Be sure to maintain updating the immediate as you employ it.
Use Claude with Claude Code, or any AI with file entry, to construct and keep the wiki. Begin at your index file when querying. Let the agent navigate.
The Sensible Challenges (And Learn how to Deal with Them)
Let’s be practical, Developing an LLM powered wiki is not any simple activity because it comes with a number of obstacles as effectively:
- Constructing an LLM-powered wiki is troublesome: It includes a number of challenges throughout setup, construction, and long-term upkeep.
- Immediate engineering is the primary problem: You want clear directions for structuring pages, deciding when to create vs replace them, and resolving conflicting data, which requires iteration and refinement.
- Scalability is a hidden issue: Easy setups break down past a number of hundred pages, so that you want tagging, folders, and search techniques deliberate prematurely.
- Consistency over time issues: With out common upkeep, your wiki will accumulate outdated data, contradictions, and orphaned pages.
- Agent proficiency is a key talent: Successfully guiding AI by prompts and construction takes observe, and people who put money into studying this get considerably higher outcomes.
Conclusion
An important recommendation for constructing your first LLM wiki is similar recommendation Karpathy offers in his gist: don’t overthink the setup. The schema template from this information might be simply copied after which you’ll be able to create the listing construction by executing the bash instructions.
The system achieves its magical impact by a number of architectural enhancements which develop from the primary day onwards. The wiki turns into extra precious with every new supply materials you embrace. The information belongs to you. The information exist in codecs which can be utilized by any system. You should use any AI you wish to question it. The LLM takes care of all upkeep duties as a substitute of you needing to deal with them which creates a unique expertise from different productiveness instruments.
Your data, lastly, working as laborious for you as you labored to accumulate it.
Continuously Requested Questions
Q1. What drawback does RAG have with data retention?
A. It doesn’t accumulate data; each question begins from scratch with out constructing on previous insights.
Q2. How does an LLM wiki differ from conventional RAG?
A. It processes data throughout ingestion, making a persistent, structured system that evolves over time.
Q3. Why is classifying paperwork earlier than ingestion necessary?
A. It ensures the system extracts the fitting stage of element for every doc kind, enhancing accuracy and usefulness.
Knowledge Science Trainee at Analytics Vidhya
I’m at the moment working as a Knowledge Science Trainee at Analytics Vidhya, the place I concentrate on constructing data-driven options and making use of AI/ML methods to unravel real-world enterprise issues. My work permits me to discover superior analytics, machine studying, and AI purposes that empower organizations to make smarter, evidence-based choices.
With a powerful basis in laptop science, software program improvement, and knowledge analytics, I’m obsessed with leveraging AI to create impactful, scalable options that bridge the hole between expertise and enterprise.
📩 You may also attain out to me at [email protected]
Login to proceed studying and revel in expert-curated content material.
Maintain Studying for Free
