okforge 0.8.0

okforge: a local-first LLM knowledge-base engine

okforge

A local-first LLM knowledge-base engine. Point it at your documents; it builds an interlinked wiki — per-document summaries, cross-document concept and entity pages, extracted images, and real page citations back to source. The wiki is plain Markdown with YAML frontmatter, readable in Obsidian or any editor, and queryable from a CLI, a chat REPL, or any MCP client.

Why

Most retrieval setups hand an LLM a pile of raw chunks and hope. okforge instead compiles your sources into curated pages ahead of time — concepts and entities that already synthesize what's spread across many documents, each claim traceable back to a (p. N) citation in the original source. That matters most for models with limited context, including small models running entirely on your own hardware: they don't have to reconstruct an answer from scratch every query, and what they do say is checkable against a specific page, not just plausible-sounding.

The output follows the Open Knowledge Format (OKF) — typed frontmatter, relative links, a predictable directory layout — so the wiki a KB produces is portable, not locked to okforge itself.

Install

pip install git+https://github.com/okforge/okforge@main

Quick start

mkdir my-kb && cd my-kb
okforge init          # scaffold raw/, wiki/, .okforge/  (--json for scripts)
okforge add paper.md  # ingest (pre-convert non-md/pdf inputs first)
okforge query "What does the paper conclude?"
okforge chat          # interactive REPL over the wiki
okforge list --json   # machine-readable state (also: status, okf-lint)
okforge describe "One line about this project."   # curated description

The query agent reads curated pages first, then drills for detail with a built-in grep_wiki lexical search (locate-then-read) rather than re-embedding everything. okf-lint checks a wiki bundle's OKF conformance.

Configuration lives in .okforge/config.yaml (model, language, entity types, …) and ~/.config/okforge/global.yaml (KB registry, default KB). The LLM endpoint is configured litellm-style — any OpenAI-compatible server works, including a local llama.cpp instance.

Ingesting scans and non-text documents

okforge add accepts Markdown, plain text, and PDF directly. Anything else — docx, pptx, scanned pages, photo catalogs — needs converting to Markdown first, by a tool that knows your material. For scanned pages specifically, okforge-vision-ocr (pip install okforge-vision-ocr) is built for exactly this: one vision-LLM call per page produces both a clean Markdown transcription and extracted photos/figures, plus a sibling <doc>.pages.json page array that okforge add reads directly for real (p. N) citations in the compiled summaries — no separate wiring needed.

okforge-vision-ocr scanned.pdf raw/book.md   # OCR + photo extraction
okforge add raw/book.md                      # ingest, with page citations

It works against any OpenAI-compatible vision-language model (tuned against a locally-hosted Qwen3.6-27B-MTP, but not tied to it).

Topic tree (experimental, per-KB opt-in)

For knowledge bases that outgrow a flat concept list: set topic_tree: true in .okforge/config.yaml, then run okforge reindex. Existing concepts cluster into named concepts/<topic>/ directories, each with a _topic.md summary node; later ingests place new concepts by tree descent, and queries gain a read_topic navigation tool for browsing top-down instead of scanning a flat list.

Wiki layout

wiki/
  index.md              # document + concept index
  summaries/<doc>.md    # per-document summary (page citations when available)
  concepts/<name>.md    # cross-document concept pages
  entities/<name>.md    # named people/places/organizations/works
  sources/<doc>.md      # ingested source text
  sources/<doc>.json    # per-page text + images (when page-aware)
  sources/images/<doc>/ # extracted images
  log.md                # append-only ingest log

Development

uv run --extra dev python -m pytest tests/   # test suite
uv run --extra dev ruff check okforge tests  # lint
uv run --extra dev ruff format okforge tests # format

Origins

okforge began as a hard fork of VectifyAI/OpenKB, in the spirit of Karpathy's LLM-wiki idea, and has since diverged deliberately rather than tracking it — local-only by default, its own document-conversion boundary, and OKF conformance as a first-class goal rather than an incidental format.

License

Apache-2.0. Portions originate from the upstream OpenKB project (copyright the original authors); okforge-specific changes are maintained at okforge/okforge.