okf-schema 0.5.0

CLI tool and Python library for working with OKF (Open Knowledge Format) bundles

okf-schema

okf-schema is a CLI tool and Python library for working with OKF (Open Knowledge Format) bundles with JSONSchema validation of the frontmatter metadata, and formatting capabilities while preserving comments.

OKF is a markdown-based knowledge format where each concept is a markdown file with YAML frontmatter. See the OKF specification for the full format definition.

📖 Full documentation: okf-schema.readthedocs.io

[!IMPORTANT] OKF-schema is opinionated. It delivers a valid OKF bundle but is adds a structure on the frontmatter that is not allowed in OKF specification:

Type values are **not** registered centrally. Producers SHOULD pick
values that are descriptive and self-explanatory; consumers MUST
tolerate unknown types gracefully (typically by treating them as
generic concepts).

In a strict OKF bundle, the type field is mandatory but can take any value and the validator needs to allow any field in the frontmatter.

OKF-schema requires the type to be one of the registered types in the _schema/ directory and validates the frontmatter against the corresponding schema. Additional properties may or may not be allowed depending on the schema definition.

What okf-schema adds to OKF

Plain OKF only defines a folder of markdown files. okf-schema turns those files into a validated, queryable knowledge base by adding:

Capability	What it does
Schema-driven frontmatter validation	Every concept's YAML frontmatter is checked against a JSONSchema. Invalid fields, missing required keys, or wrong types are reported as structured errors.
Auto-discovered schemas	Schemas live inside the bundle under _schema/ (e.g. _schema/concept.schema.yaml). The type field in a concept's frontmatter tells okf-schema which schema file to load. A concept with type: concept is validated against _schema/concept.schema.yaml. Schemas can be written in YAML, JSON, or JSON5 (JSON with comments and trailing commas).
Bundle integrity checks	Detects broken internal links, missing index.md files, malformed log.md entries, and reserved-file violations.
Safe linting	Normalizes YAML frontmatter by flattening nested lists and converting block-style to inline notation while preserving comments and custom quotes via ruamel.yaml. Also auto-updates links and backlinks fields from markdown body content.
Analytics	Bundle statistics.

See a real schema definition in examples/ai-llm-knowledge-base/_schema/concept.schema.yaml.

Example of structure

my-bundle/
├── _schema/
│   ├── concept.schema.yaml
│   ├── tool.schema.json
│   └── paper.schema.json5
├── concepts/
│   ├── rag.md
│   └── chain-of-thought.md
├── tools/
│   ├── langchain.md
│   └── llamaindex.md
├── papers/
│   ├── rag-paper.md
│   └── chain-of-thought-paper.md
├── index.md
└── log.md

The type field in each entity frontmatter determines which schema is used for validation. For example, type: concept uses _schema/concept.schema.yaml, while type: tool uses _schema/tool.schema.json.

Schema extensions supported:

• .schema.yaml — YAML (human-friendly, supports comments and anchors)
• .schema.json — JSON (strict syntax, widely supported by editors)
• .schema.json5 — JSON5 (JSON with comments, trailing commas, and unquoted keys)

$ref support

Schemas can reference external files with $ref. The path is resolved relative to the _schema/ directory:

# _base.schema.yaml
$schema: "https://json-schema.org/draft/2020-12/schema"
type: object
properties:
  type:
    type: string
  title:
    type: string
required:
  - type
additionalProperties: true

# concept.schema.yaml
$ref: _base.schema.yaml
properties:
  category:
    enum: [LLM, AI Agent, Coding Agent]

$ref works at any nesting level (top-level, inside properties, inside items, etc.). If the referenced file cannot be found, the $ref is preserved as-is and validation proceeds with the remaining schema content.

Default _base.schema.yaml

When you run okf-schema init, a _base.schema.yaml is automatically created in _schema/. It documents the standard OKF frontmatter fields and can be used as a $ref target for your own schemas:

Field	Required	Description
type	Yes	A short string identifying the kind of concept.
title	No	Human-readable display name.
description	No	One-line summary of the concept.
resource	No	URI of the underlying asset.
tags	No	List of short categorization strings.
timestamp	No	ISO 8601 datetime of last change.

Installation

uv tool install okf-schema

Quickstart

# Initialize a new OKF bundle
okf-schema init my-bundle

# Update index.md files for all directories
okf-schema index --path my-bundle/bundle

# Lint frontmatter (flatten nested lists, inline block-style, auto-update links/backlinks)
okf-schema lint --path my-bundle/bundle

# Lint without updating links/backlinks
okf-schema lint --path my-bundle/bundle --no-links

# Validate a bundle
okf-schema validate --path my-bundle/bundle
# or enforce strict validation (fail on warnings)
okf-schema validate --path my-bundle/bundle --strict

# List all concepts
okf-schema list --path my-bundle/bundle

# Find backlinks to a concept (extension is optional)
okf-schema backlinks --path my-bundle/bundle concepts/react-pattern

CLI Reference

Subcommand	Description
init <name>	Create a new OKF bundle directory structure
new --path <dir> --name <name>	Create a new concept file with frontmatter template
validate --path <bundle>	Validate bundle structure and frontmatter
validate --path <bundle> --strict	Validate and fail on warnings
lint --path <bundle>	Lint frontmatter: flatten nested lists, convert block-style to inline, and auto-update links/backlinks from markdown body
list --path <bundle>	List all concepts in a bundle
show --path <bundle> <concept>	Show a single concept's frontmatter and body
index --path <bundle>	Regenerate all index.md files
stats --path <bundle>	Show bundle statistics
backlinks --path <bundle> <target>...	List concepts that link to the given target(s)

Recommended Workflow

Before packaging or distributing a bundle, run these three commands in order and fix all warnings:

okf-schema index --path my-bundle/bundle    # regenerate index.md files
okf-schema lint --path my-bundle/bundle     # flatten nested lists, inline block lists, update links/backlinks
okf-schema validate --path my-bundle/bundle --strict # check structure, schema, and links; fail on warnings

Only zip or ship the bundle once validate --strict reports zero errors and zero warnings. Warnings such as missing index.md (W4), block-style lists (W7), or broken cross-links (W2) signal issues that will degrade the experience for downstream consumers.

Example: AI & LLM Knowledge Base

The examples/ai-llm-knowledge-base/ directory contains a realistic knowledge base with three concept types — concept, tool, and paper — each validated by its own schema in _schema/.

How type selects the schema

The type field in a concept's frontmatter determines which schema file is loaded. A file with type: concept is validated against _schema/concept.schema.yaml; type: tool against _schema/tool.schema.json; and type: paper against _schema/paper.schema.json5.

Schema format support

okf-schema accepts schemas in three formats:

Extension	Format	Notes
.schema.yaml	YAML	Human-friendly, supports comments and anchors
.schema.json	JSON	Strict syntax, widely supported by editors
.schema.json5	JSON5	JSON with comments, trailing commas, and unquoted keys

Schema highlights

concept.schema.yaml — AI concepts with enums, email validation, and kebab-case regex:

properties:
  category:
    enum: [LLM, AI Agent, Coding Agent, Prompt Engineering, Tooling, Evaluation]
  maturity:
    enum: [experimental, beta, production, deprecated]
  author_email:
    type: string
    format: email
  tags:
    type: array
    items:
      pattern: "^[a-z0-9-]+$"   # kebab-case only

tool.schema.json — Developer tools with URI validation and language enums:

{
  "properties": {
    "license": {
      "enum": ["MIT", "Apache-2.0", "GPL-3.0", "Proprietary", "Other"]
    },
    "language": {
      "enum": ["Python", "JavaScript", "TypeScript", "Rust", "Go", "Java", "Multi-language"]
    },
    "url": { "type": "string", "format": "uri" }
  }
}

paper.schema.json5 — Research papers with year bounds and venue enums:

// JSON5 allows comments, trailing commas, and unquoted keys
{
  properties: {
    year: { type: "integer", minimum: 1950, maximum: 2030 },
    venue: {
      enum: ["NeurIPS", "ICML", "ICLR", "ACL", "EMNLP", "arXiv", "Other"]
    },
    bibtex_key: { pattern: "^[A-Za-z0-9_-]+$" },
  },
}

Schema-aware index generation

Schemas can declare a title and an x-okf-summary extension field. When a subdirectory contains concepts of a single type, okf-schema index uses these values to produce richer index.md files:

Field	Purpose	Used in
title	Short heading for the concept type	Subdirectory index.md H1
x-okf-summary	One-line description of the type	Subdirectory intro + root listing
description	Fallback when x-okf-summary is absent	Same places as above

For example, concept.schema.yaml declares:

title: "Concept"
x-okf-summary: "AI/LLM concepts such as techniques, patterns, or architectural ideas."
description: "Schema for AI/LLM concepts ..."

Running okf-schema index turns this into:

• A root index.md entry: [concepts](./concepts/) — AI/LLM concepts such as...
• A subdirectory index.md with # Concept as the heading and the summary as the first paragraph.

Concept file example (concepts/rag.md)

---
type: concept
title: Retrieval-Augmented Generation
description: >
  A technique that enhances LLM outputs by retrieving relevant documents
  from an external knowledge store and injecting them into the prompt.
category: LLM
maturity: production
author_email: [e-mail]
complexity: intermediate
tags: [rag, retrieval, llm, knowledge-base]
related_tools: [LangChain, LlamaIndex, OpenAI-API]
---

# Retrieval-Augmented Generation

RAG combines parametric knowledge (the model's weights) with non-parametric
knowledge (external documents) to reduce hallucinations...

Validation in action

# Validates all concepts, tools, and papers against their respective schemas
okf-schema validate --path examples/ai-llm-knowledge-base

# Show bundle statistics
okf-schema stats --path examples/ai-llm-knowledge-base

Python API

from okf_schema.api import validate_bundle

report = validate_bundle("path/to/bundle")
for finding in report.findings:
    print(finding.level, finding.message)

# The _schema/ directory inside the bundle is auto-discovered.
# You can also pass an explicit schema_db path:
# report = validate_bundle("path/to/bundle", schema_db="path/to/schemas")

Documentation

Documentation is available at okf-schema/README.md.

Evaluation

okf-schema includes an automated skill evaluation campaign in skills-evals/ that validates the tool against a suite of fixture bundles. These evals verify that the validator correctly identifies conformant bundles, catches structural errors (E1–E6), reports warnings (W1–W7), and handles JSONSchema validation with both JSON and YAML schema databases.

Eval	Description
validate-conformant-bundle	Validates a fully conformant OKF bundle (0 errors, 0 warnings)
validate-non-conformant-bundle	Validates a bundle with known errors (E1–E3) and warnings (W1–W7)
create-okf-bundle	Creates a new OKF bundle from scratch and validates it
validate-with-schema-database	Tests JSONSchema validation with --schema-db
validate-with-yaml-schema-database	Tests YAML schema file support in schema DB

Run the eval campaign by asking in Copilot Chat Execute instructions in skills-evals/eval.prompt.md, or via Copilot-CLI:

# Trigger eval via Copilot-CLI
just copilot-cli-eval-okf-schema
# Open eval review in browser
just eval-view-okf-schema

The eval system supports A/B comparison (with_skill vs without_skill) across iteration directories. Results are rendered as an interactive HTML dashboard.

Latest eval result: iteration-30.06.26-22.06 — skeptical review

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT License — see LICENSE for details.