OpenKB — Open LLM Knowledge Base

Scale to long documents  •  Reasoning-based retrieval  •  Native multi-modality  •  No Vector DB

---

📑 What is OpenKB

OpenKB (Open Knowledge Base) is an open-source system (in CLI) that compiles raw documents into a structured, interlinked wiki-style knowledge base using LLMs, powered by PageIndex for vectorless long document retrieval.

The idea is based on a concept described by Andrej Karpathy: LLMs generate summaries, concept pages, and cross-references, all maintained automatically. Knowledge compounds over time instead of being re-derived on every query.

Why not just traditional RAG?

Traditional RAG rediscovers knowledge from scratch on every query. Nothing accumulates. OpenKB compiles knowledge once into a persistent wiki, then keeps it current. Cross-references already exist. Contradictions are flagged. Synthesis reflects everything consumed.

Features

• Any format — PDF, Word, PowerPoint, Excel, HTML, Markdown, text, CSV, and more via markitdown
• Scale to long documents — Long and complex documents are handled via PageIndex tree indexing, enabling accurate, vectorless long-context retrieval
• Native multi-modality — Retrieves and understands figures, tables, and images, not just text
• Auto wiki — LLM generates summaries, concept pages, and cross-links. You curate sources; the LLM does the rest
• Query — Ask questions against your wiki. The LLM navigates your compiled knowledge to answer
• Lint — Health checks find contradictions, gaps, orphans, and stale content
• Watch mode — Drop files into raw/, wiki updates automatically
• Obsidian compatible — Wiki is plain .md files with [[wikilinks]]. Open in Obsidian for graph view and browsing

🚀 Getting Started

Install

pip install openkb

Quick start

# 1. Create a directory for your knowledge base
mkdir my-kb && cd my-kb

# 2. Initialize the knowledge base
openkb init

# 3. Add documents
openkb add paper.pdf
openkb add ~/papers/                   # Add a whole directory
openkb add article.html

# 4. Ask questions
openkb query "What are the main findings?"

# 5. Check wiki health
openkb lint

Set up your LLM

OpenKB comes with multi-LLM support (e.g., OpenAI, Claude, Gemini) via LiteLLM (pinned to a safe version).

Set your model during openkb init, or in .openkb/config.yaml, using provider/model LiteLLM format (like anthropic/claude-sonnet-4-6). OpenAI models can omit the prefix (like gpt-5.4).

Create a .env file with your LLM API key:

LLM_API_KEY=your_llm_api_key

🧩 How It Works

Architecture

raw/                              You drop files here
 │
 ├─ Short docs ──→ markitdown ──→ LLM reads full text
 │                                     │
 ├─ Long PDFs ──→ PageIndex ────→ LLM reads document trees
 │                                     │
 │                                     ▼
 │                         Wiki Compilation (using LLM)
 │                                     │
 ▼                                     ▼
wiki/
 ├── index.md            Knowledge base overview
 ├── log.md              Operations timeline
 ├── AGENTS.md           Wiki schema (LLM instructions)
 ├── sources/            Full-text conversions
 ├── summaries/          Per-document summaries
 ├── concepts/           Cross-document synthesis ← the good stuff
 ├── explorations/       Saved query results
 └── reports/            Lint reports

Short vs. long document handling

	Short documents	Long documents (PDF ≥ 20 pages)
Convert	markitdown → Markdown	PageIndex → tree index + summaries
Images	Extracted inline (pymupdf)	Extracted by PageIndex
LLM reads	Full text	Document trees
Result	summary + concepts	summary + concepts

Short docs are read in full by the LLM. Long PDFs are indexed by PageIndex into a hierarchical tree with summaries. The LLM reads the tree instead of the full text, enabling better retrieval from long documents.

The wiki compiles knowledge

When you add a document, the LLM:

1. Generates a summary page
2. Reads existing concept pages
3. Creates or updates concepts with cross-document synthesis
4. Updates the index and log

A single source might touch 10-15 wiki pages. Knowledge accumulates: each document enriches the existing wiki rather than sitting in isolation.

📦 Usage

Commands

Command	Description
openkb init	Initialize a new knowledge base (interactive)
openkb add <file_or_dir>	Add documents and compile to wiki
openkb query "question"	Ask a question against the knowledge base
openkb query "question" --save	Ask and save the answer to wiki/explorations/
openkb watch	Watch raw/ and auto-compile new files
openkb lint	Run structural + knowledge health checks
openkb list	List indexed documents and concepts
openkb status	Show knowledge base stats

Configuration

Settings are initialized by openkb init, and stored in .openkb/config.yaml:

model: gpt-5.4                   # LLM model (any LiteLLM-supported provider)
language: en                     # Wiki output language
pageindex_threshold: 20          # PDF pages threshold for PageIndex

Model names use provider/model LiteLLM format (OpenAI models can omit the prefix):

Provider	Model example
OpenAI	gpt-5.4
Anthropic	anthropic/claude-sonnet-4-6
Gemini	gemini/gemini-3.1-pro-preview

PageIndex integration

Long documents are challenging for LLMs due to context limits, context rot, and summarization loss. PageIndex solves this with vectorless, reasoning-based retrieval — building a hierarchical tree index that lets LLMs reason over the index for context-aware retrieval.

PageIndex runs locally by default using the open-source version, with no external dependencies required.

Optional: Cloud Support

For large or complex PDFs, PageIndex Cloud can be used to access additional capabilities, including:

• OCR support for scanned PDFs (via hosted VLM models)
• Faster structure generation
• Scalable indexing for large documents

Set PAGEINDEX_API_KEY in your .env to enable cloud features:

PAGEINDEX_API_KEY=your_pageindex_api_key

AGENTS.md

The wiki/AGENTS.md file defines wiki structure and conventions. It's the LLM's instruction manual for maintaining the wiki. Customize it to change how your wiki is organized.

At runtime, the LLM reads AGENTS.md from disk, so your edits take effect immediately.

Using with Obsidian

OpenKB's wiki is a directory of Markdown files with [[wikilinks]]. Obsidian renders it natively.

1. Open wiki/ as an Obsidian vault
2. Browse summaries, concepts, and explorations
3. Use graph view to see knowledge connections
4. Use Obsidian Web Clipper to add web articles to raw/

🧭 Learn More

Compared to Karpathy's Approach

	Karpathy's workflow	OpenKB
Short documents	LLM reads directly	markitdown → LLM reads
Long documents	Context limits, context rot	PageIndex tree index
Supported formats	Web clipper → .md	PDF, Word, PPT, Excel, HTML, text, CSV, .md
Wiki compilation	LLM agent	LLM agent (same)
Q&A	Query over wiki	Wiki + PageIndex retrieval

Tech Stack

• PageIndex — Vectorless, reasoning-based document indexing and retrieval
• markitdown — Universal file-to-markdown conversion
• OpenAI Agents SDK — Agent framework (supports non-OpenAI models via LiteLLM)
• LiteLLM — Multi-provider LLM gateway
• Click — CLI framework
• watchdog — Filesystem monitoring

Roadmap

• Extend long document handling to non-PDF formats
• Scale to large document collections with nested folder support
• Hierarchical concept (topic) indexing for massive knowledge bases
• Database-backed storage engine
• Web UI for browsing and managing wikis

Contributing

Contributions are welcome! Please submit a pull request, or open an issue for bugs or feature requests. For larger changes, consider opening an issue first to discuss the approach.

License

Apache 2.0. See LICENSE.

Support Us

If you find OpenKB useful, give us a star 🌟 — and check out PageIndex too!