LLM Knowledge Base

Built by Artur Ferreira @ The GEO Lab · 𝕏 @TheGEO_Lab · LinkedIn · Reddit

A schema and workflow for LLM-compiled personal knowledge bases.

Inspired by Andrej Karpathy's April 2026 post on using LLMs to build and maintain markdown wikis, this repo formalises the pattern into a reusable, versioned schema — including a structured learning layer that the original workflow leaves out.

The core idea: you collect raw source material. The LLM compiles it into a structured wiki. You never edit the wiki manually. You just learn from it.

What this repo contains

The workflow in one diagram

raw/                          wiki/                        output/
───────────────────           ──────────────────────       ──────────────────
articles/          →  LLM  →  _index.md                →  reports/
papers/            compile    _concepts.md              →  slides/  (Marp)
repos/                        _graph.md                 →  figures/ (matplotlib)
datasets/                     concepts/*.md
images/                       summaries/*.md       ↑
                              topics/*.md          └── filed back into wiki
                                        ↓
                              learning/
                              _review.md           ← spaced repetition queue
                              flashcards/*.md      ← LLM-generated Q&A
                              gaps.md              ← open questions

Quickstart

1. Clone and initialise your vault

git clone https://github.com/arturseo-geo/llm-knowledge-base
cd llm-knowledge-base
cp -r templates/ my-wiki/

Or just copy AGENTS.md into the root of an existing Obsidian vault.

2. Drop source material into raw/

Anything works: PDFs, markdown files clipped with Obsidian Web Clipper, .py scripts, .csv datasets, images. The LLM handles all formats.

3. Point your LLM at the repo and say:

Read AGENTS.md, then compile everything in raw/ into the wiki.

That's it. The agent reads the schema, creates the index files, writes concept articles, summaries, backlinks, and flashcards. From that point on, incremental ingest is a single instruction:

File this to our wiki: raw/papers/new-paper.pdf

4. Query against the wiki

What are the key differences between RLHF and Constitutional AI?
Write the answer as a report and save to output/reports/.

5. Review your learning queue

Show me what's due for review in learning/_review.md and run a flashcard session.

The learning layer

The gap in every existing LLM knowledge base workflow: collecting and organising knowledge is not the same as mastering it.

This schema adds a structured learning layer under learning/:

• Flashcards are generated automatically whenever a concept article is created or updated
• A review queue tracks due dates using a simplified FSRS spaced repetition algorithm
• Gap tracking captures open questions the LLM detects during linting — turning weak spots in your wiki into a research agenda
• Socratic Q&A — ask the LLM to quiz you on any concept, and it grades answers against the wiki, not against general knowledge

The result: your wiki compounds twice. Once as a knowledge store, and again as a learning system that actively surfaces what you don't yet know.

AGENTS.md — the schema spec

AGENTS.md is the contract between you and any LLM operating on your knowledge base. It defines:

• Directory layout and file naming conventions
• Frontmatter schema for every wiki article
• How the three index files (_index.md, _concepts.md, _graph.md) are maintained
• Step-by-step compilation, query, and linting workflows
• Learning layer: flashcard format, scheduling rules, gap tracking
• Quality rules: no hallucinated citations, confidence levels, orphan prevention
• Contamination mitigation: sandbox-first promotion model
• Session startup checklist

The spec is versioned. Breaking changes increment the major version. The agent checks the version header on every session and adapts to schema changes immediately.

Read the full spec →

Why not RAG?

For personal and team knowledge bases at typical research scale (50–1,000 articles), a compiled markdown wiki outperforms a vector RAG stack on every practical dimension:

At scale (millions of documents, heterogeneous corpora, sub-second latency requirements), RAG is the right call. For a focused research domain, you probably don't need a million documents — you need the right 200, well-compiled and well-linked.

Recommended tools

IDE: Obsidian — local-first markdown vault with graph view, backlinks, and a plugin ecosystem. Free for personal use.

Ingest: Obsidian Web Clipper — clips web articles to .md directly into your vault.

LLM: Claude (Sonnet for compilation and Q&A, Haiku for cheap linting sweeps), or any model with a large context window and good instruction-following.

Slides: Marp — render Marp-format markdown as presentation slides, viewable in Obsidian.

Search: Start with ripgrep over the wiki directory. Upgrade to a local vector index (ChromaDB + Nomic Embed via Ollama) only when you exceed ~1,000 articles.

Version control: Git. Your wiki is just files — version it like code.

Contamination mitigation

Agent-generated content carries quality risk — and in Obsidian, the risk goes beyond content quality. Obsidian's search, graph, backlinks, and quick switcher all operate at the vault level. Once agent-compiled articles mix with your personal notes, Obsidian stops being a representation of your knowledge and becomes a representation of the LLM's synthesis.

The primary solution is the two-vault model: one Obsidian vault for the agent (full schema), one for your personal notes (insights/ — human-written only, agent never touches it). See docs/two-vault-setup.md for the full setup guide.

For single-repo setups, the schema enforces a sandbox-first model:

• The LLM writes freely to output/ and learning/
• Promotion to wiki/ requires either explicit instruction or passing the quality rules in AGENTS.md §11
• The agent never writes to insights/ — this directory is human-only by schema rule
• Articles with confidence: speculative are not linked into the graph until upgraded
• The linting workflow quarantines contradictory articles with a status: quarantined flag

The distinction Steph Ango (@kepano) put precisely: a summary of a PDF is noise. An insight you formed from reading it is signal. The schema keeps them in separate directories with separate authorship rules.

Finetune path

As your wiki grows and is repeatedly linted, it becomes a progressively cleaner, deduplicated, consistently-styled representation of your research domain. At that point it is more than a context store — it is a candidate training corpus.

The learning/flashcards/ directory generates structured Q&A pairs automatically. Combined with the concept articles and summaries, this gives you the raw material for supervised fine-tuning of a smaller, domain-specialised model.

Tools for this path: Distilabel for synthetic data generation, Unsloth or Axolotl for LoRA fine-tuning.

Full finetune guide →

Contributing

The schema is intentionally opinionated. If you use it and find patterns that work better, open an issue or PR. The goal is a community-maintained standard, not a personal config file.

Areas where contributions are most welcome:

• Additional output format templates (reveal.js, Jupyter notebooks)
• Alternative scheduling algorithms for the learning layer
• Worked example wikis on different research domains
• CLI tooling that automates the linting workflow

See CONTRIBUTING.md for guidelines.

Schema version

Current: 1.1.0 — see CHANGELOG.md for full history.

Related

• Andrej Karpathy's original post — the workflow that inspired this
• Karpathy's follow-up gist — full architecture, philosophy, and tooling
• Steph Ango on contamination and vault separation — the two-vault model and insight vs noise distinction, by Obsidian's co-creator
• Obsidian Spaced Repetition plugin — manual flashcards in Obsidian; this schema automates generation
• Project Second Brain — a more complex implementation with Neo4j and a web app
• awesome-llm-knowledge-bases — curated tool list for the ecosystem

Attributions & Licence

Built and maintained by Artur Ferreira · Part of the The GEO Lab toolkit.

MIT — see LICENSE. Copy freely, adapt for your domain, keep the schema versioned.

Found this useful? ⭐ Star the repo and connect: 🌐 thegeolab.net · 𝕏 @TheGEO_Lab · LinkedIn · Reddit

Related Repos

• claude-code-skills — All 17 skills in one collection
• seo-geo-skill — SEO & GEO optimisation skill
• grounded-research-skill — Anti-hallucination research mode
• content-pipeline-skill — 7-agent content pipeline with GEO Quality Gate