Knowledge Graph for Education Global

A configurable, provenance-preserving pipeline for transforming curriculum PDFs into validated knowledge graphs aligned to the Learning Commons ontology.

The pipeline separates document reconstruction from curriculum-semantic interpretation so that intermediate decisions can be inspected, validated, and audited.

Pipeline

Curriculum PDF
    |
    v
Page IR Extraction
    |
    v
Page IR Verification
    |
    v
Document IR
    |
    v
Academic Standards KG
    |
    v
Academic Standards + Learning Components KG

The five conceptual stages are implemented through four CLI entry points. The final entry point builds the validated Academic Standards layer first and then derives Learning Components from it.

Quick start

Follow the local setup guide, create or adapt a runtime config, and run the pipeline from backend/:

python src/kgfeg/entries/extract_page_ir.py <config.json>
python src/kgfeg/entries/verify_page_ir_continuity.py <config.json>
python src/kgfeg/entries/stitch_document_ir.py <config.json>
python src/kgfeg/entries/create_kgs.py <config.json>

Example curriculum profiles are available under examples/. If you are adapting the system to a new source, start with the Add a New Curriculum guide rather than editing the backend for source-specific conventions.

What the pipeline produces

The run preserves intermediate evidence for debugging and auditability while producing two main validated KG handoffs:

• kgs/as_kg_bundle.json — Academic Standards framework, items, hierarchy, and provenance; and
• kgs/as_lc_kg_bundle.json — the combined Academic Standards + Learning Components graph with hasChild and supports relationships.

See the pipeline overview for the full artifact map and stage contracts. Downstream consumers should also read the output artifacts and integration contract before building against the exported JSON/JSONL shapes.

Documentation

• Official documentation
• Architecture
• Pipeline overview
• Add a new curriculum
• Run, resume, and debug
• Output artifacts and integration contract
• Local development setup
• Contributing

Design principles

• Source grounded: final standards and skills remain traceable to source evidence.
• Stage separated: PageIR and DocumentIR reconstruct the document before curriculum semantics are inferred.
• Validated: higher-risk LLM assertions use explicit deterministic checks and, where appropriate, independent producer/checker flows.
• Deterministic where possible: identity construction, graph constraints, reconciliation checks, and final compilation are enforced in Python.
• Configuration driven: curriculum-specific taxonomy, hierarchy, codes, extraction policy, and LC policy belong in document profiles rather than source-specific backend branches.

Contact

See Contact us for project and team contact information.