ski-kg-extractor 3.0.3

Extract compliance rules from regulatory documents using LLM-assisted parsing. Phase 1 only.

kg-extractor

Install from PyPI: pip install ski-kg-extractor (publishing starts with the first release after June 2026).

Status: v3.0 — first production-target release.

Extract structured compliance rules from regulatory documents (LLM-assisted) and emit a v3 Knowledge Graph ready for kg-validator and the SKI Model runtime.

kg-extractor is a Phase 1 (compilation) tool. The output goes through kg-validator (schema + §3.6 validation passes) and is then signed before crossing the sovereign boundary. The extractor's LLM backend may be local or cloud — its output is not part of the runtime path.

Highlights (v3.0)

• Emits v3 KG JSON by default. Each extracted rule is wrapped into the typed-graph shape from spec v3.0 §3: a Rule node, an Obligation with a guessed obligation_type and numeric value/unit, a Subject, a Citation, and the four edges connecting them (applies_to, consists_of, scoped_to, cited_by).
• --jurisdiction flag. Every KG carries a Jurisdiction node and every rule is scoped_to it. Defaults to global.
• Pluggable LLM backend. Default is Ollama (local). Cloud backends (anthropic, openai) exist for compilation-phase use only.
• Deterministic by default. Temperature 0, fixed seed, recorded prompt SHA-256 in extraction metadata for reproducibility audits.
• IMPLIED is prohibited (B2.1 Anchor Constraint). The prompt instructs the model not to emit IMPLIED and the parser downgrades any that slip through to DISCRETIONARY with a warning. Rules cannot be inferred beyond source text.
• extraction_quality replaces confidence. The per-rule trust signal is now extraction_quality (EXPLICIT / DISCRETIONARY / CONFLICTING) — a Phase-1 authoring concept, separate from the runtime's categorical verdict (Axiom 2 prohibits confidence scores in the audit trail).

Installation

pip install -e tools/kg-extractor

Quick start

# Extract a v3 KG from a regulation:
kg-extractor extract --file regulation.txt --sector energy \
    --jurisdiction us.federal \
    --output kg-extracted-v3.json

# Validate the emitted KG:
kg-validator validate --input kg-extracted-v3.json

# Batch:
kg-extractor batch --input-dir ./regulations --output-dir ./kgs \
    --sector energy --jurisdiction us.federal

# Debugging — dump the raw flat-rule list before wrapping:
kg-extractor extract --file regulation.txt --emit-raw --output flat.json

Architecture

kg_extractor/
├── __init__.py     public API: Extractor, ExtractionResult, ExtractionQuality, emit_v3_kg
├── extractor.py    chunk → LLM → flat ComplianceRule list
├── models.py       ComplianceRule, ExtractionResult, ExtractionQuality enum
├── backends.py     Ollama / Anthropic / OpenAI backends (compilation phase only)
├── v3_emitter.py   wraps ExtractionResult into a v3 KG dict (spec §3 shape)
├── utils.py        chunking, document parsing, rule validation
└── cli.py          click-based CLI (`extract`, `batch`, `filter`, `examples`, `version`)

Notes on extraction_quality

• EXPLICIT — the LLM produced a verbatim source quote in reasoning. The rule is safe to ship without further review.
• DISCRETIONARY — the LLM could not anchor the rule in a verbatim source quote. A human reviewer must approve before shipping.
• CONFLICTING — the LLM detected internal contradiction in the source. Surface for legal review.

This is not a confidence score in the probabilistic sense. The runtime never reads it. It exists solely to drive the Phase-1 human-review queue.