Your Obsidian vault has 2,000 notes. Your research folder has 40 PDFs. When you need a synthesized answer from that material, you still end up digging through folders and skimming documents to reconstruct the context yourself. The information exists, but turning it into usable knowledge still depends on you.
Andrej Karpathy described this as the LLM Wiki problem. His proposed solution reframes how AI interacts with documents: an LLM reads your files incrementally and compiles them into a structured, evolving knowledge base, so the system gradually builds on what it already knows
This article shows how to build that workflow on Zo. You'll take an existing folder of notes, PDFs, or markdown files, sync them into a persistent workspace, and build a knowledge base you can query conversationally.
From Notes to a Knowledge System
Tools like Obsidian, Notion, and Google Docs are built for storing information and returning keyword matches. Search a term and you get files back. The synthesis still falls on you.
A paper you saved six months ago might connect directly to something you documented last week, but those links stay buried unless you happen to remember both sources at the same time. Even after building a large collection of notes, every query still feels isolated because search gives you files, not understanding across the collection.
Notion AI and Google NotebookLM improve this by layering AI on top of search. Retrieval gets better, but the workflow still depends on searching documents inside a hosted platform. The system surfaces information more efficiently, but it still behaves more like a smarter search engine than a knowledge layer that evolves alongside your files. .
Why Traditional RAG Workflows Feel Incomplete
Retrieval-Augmented Generation improves how language models work with external documents. Instead of relying only on training data, the model searches your files for relevant passages and uses them as context before generating a response grounded in your source material.
In a typical implementation, documents are broken into chunks, converted into embeddings, and searched for sections that are semantically similar to a query. This works well when you need answers tied back to source documents.
The limitation is that the workflow still revolves around retrieval. Each query searches raw documents again, even if the system has already processed similar material before. The interaction helps answer the question, but it does not really build persistent structure over time. Many RAG setups also rely on several moving parts: an embedding provider, a vector database, a retrieval layer, and an inference endpoint. Maintaining that pipeline can easily become part of the project itself.
The LLM Wiki pattern shifts the focus toward continuously building structured understanding. Instead of treating documents as isolated chunks waiting to be searched, the system gradually turns them into a persistent knowledge layer that grows alongside the material you keep adding.
Zo's Approach: From Folder to Persistent Knowledge Layer
Zo runs your files, your compiled notes, and your AI interface in one persistent workspace. Your files stay available between sessions, updates remain in place, and the AI continues working against the same evolving set of documents over time. The workspace stays persistent because Zo runs on an always-on Elastic Computer environment behind the scenes
That continuity changes how the knowledge base behaves. Notes, PDFs, markdown files, and generated summaries all live in the same environment, so the system can build on existing context as your collection grows. You upload once, and the workspace keeps your material accessible across every session.
Zo's file browser and desktop sync make this feel closer to maintaining a live folder than managing a retrieval pipeline. You can sync an existing Obsidian vault, upload research PDFs, or keep markdown notes updated directly from your desktop. As files change, the workspace reflects those updates automatically.
Ingesting and Structuring Your Knowledge Base
Upload your documents through Zo's file browser. Drag in an Obsidian vault export, a folder of research PDFs, plain text notes, or any combination. Keeping raw documents, compiled notes, and inbox folders separated from the start makes the workflow easier to maintain as the collection grows:
/knowledge-base/
/sources/
/pdfs/
/markdown/
/obsidian-export/
/compiled/
/topics/
/concepts/
/connections/
/index.md
/inbox/
The /sources/ directory holds raw documents. The /compiled/ directory stores the structured markdown notes generated during the compilation process. The /inbox/ folder stages new documents before processing.
With your files uploaded, open the Zo workspace and run compilation prompts against your source documents. The parsing prompt below extracts key concepts from a PDF and writes them into a structured note under /compiled/topics/:
Read /knowledge-base/sources/pdfs/[filename].pdf and extract the key
concepts and how they relate to each other. Create a structured markdown note at
/knowledge-base/compiled/topics/[topic-name].md with the following sections:
- Topic overview (2-3 sentences)
- Key concepts, each with a one-sentence definition
- Related topics (list only topic names for cross-linking)
- Source reference (filename and relevant page numbers)
Keep formatting consistent with other notes in /knowledge-base/compiled/topics/.
For cross-linking concepts across notes, run a second pass. The cross-linking prompt below identifies concept pairs that appear across multiple notes with no explicit connection, writes a short relationship note for each, and keeps index.md current:
Review the notes in /knowledge-base/compiled/topics/ and identify up to 5
pairs of concepts that appear in multiple notes but have no explicit link
between them. For each pair, write a short connection note explaining the
relationship and save it to /knowledge-base/compiled/connections/.
Then update /knowledge-base/compiled/index.md with a full list of all topics
and their connections.
These prompts build the knowledge base incrementally, adding structure over time without forcing you to rebuild the entire workflow. When you add new sources to /inbox/, run the parsing prompt against the new files, then run the cross-linking pass to integrate them into the existing structure.
Each new document moves through the same compilation cycle before landing in the queryable compiled set:
Over time, the index becomes a practical navigation layer for everything the system has compiled.
Configuring the Knowledge and Query Layer
Zo lets you pick a model directly from the workspace UI. You can use one of the built-in options (Claude, GPT-4o, Gemini, DeepSeek) or connect your own through Zo's BYOK settings. Once you add your API key, the model appears in the dropdown, ready to use inside the same workspace as your files and compiled notes. As better models become available, you can swap them without disrupting the workflow.
From there, configure a Persona and a set of Rules through Zo's Personas interface to anchor the system's behavior to your knowledge base. A starting configuration for a research knowledge base looks like this:
Persona: Research Knowledge Base
You are a knowledge assistant for [your domain]. Answer questions using only
the notes in /knowledge-base/compiled/. For every answer, cite the specific
note or source file used. If a question falls outside the compiled knowledge,
say so clearly and identify which topic area is missing.
Formatting rules:
- Lead with a direct answer (2-3 sentences)
- Follow with supporting detail from the notes
- End with source citations (filename and topic name)
With this Persona active, the system answers from compiled notes. Coverage deltas in the knowledge base become visible quickly. When the compiled notes lack material on a topic, the system flags it and identifies which area needs more source material.
Rules can tighten the workflow further. You might require every response to include a citation, flag topics with fewer than two source notes, or instruct the system to suggest new compilation targets when a subject area is underdeveloped.
Using the LLM Wiki Day-to-Day
Once the knowledge base is in place, querying it produces synthesized answers drawn from everything you've compiled. You can ask broad research questions and get responses built from connected topic notes, each one citing back to source files.
A query like "What's the relationship between attention mechanisms and positional encoding?" returns a response grounded in the compiled topic notes, with citations pointing back to specific files. The system works from the structured layer built during compilation.
Cross-collection queries work the same way. Ask "Which papers in my corpus are most relevant to retrieval-augmented generation?" and the system draws from the index and connection notes, surfacing relevant topics across your entire compiled set.
Adding new material stays straightforward. Drop files into /inbox/, run the parsing prompt, then run the cross-linking pass to integrate new concepts into the existing structure.After several weeks of consistent additions, the knowledge base starts surfacing relationships no individual document exposes on its own.
If you want access outside the workspace, Zo supports publishing the interface through Zo Sites with a custom domain. The published interface works from the same workspace and files, so updates remain available as the knowledge base grows.
What to Build Today
If you have an Obsidian vault, a folder of PDFs, or a collection of markdown notes, you have enough to start.
Open your Zo workspace, upload documents through the file browser, and set up the folder structure above. Start small: run the parsing prompt against five to ten source files, then run the cross-linking pass to build the first set of connections.
Ask the system a real question from your current work and compare the response to what you'd find searching the raw documents yourself. Compiled notes surface relationships and context that stay invisible when information lives in separate files.
Expand from there. Add new research folders as the collection grows and refine your Personas and Rules as you learn what the system does well. The Zo tutorials library covers deeper workflows for file syncing, Persona configuration, and workspace organization.
Upload your documents and start building your own LLM Wiki at Zo Computer.
Top comments (0)