najonghwa/safety-law-db

Purpose and Deliverables

This repository builds a local SQLite, article-level search database for Korean safety and environmental regulations. It gathers laws, administrative notices, annexes, KOSHA guidance, and chemical-substance lists, then packages a Claude Code-facing search workspace. Primary target definitions are in laws.json; vocabulary normalization is in synonyms.json.

Pipeline Entry Points

The intended batch sequence is:

1. Collect legal texts: collector.py

2. Collect annexes and parse HWP/text: annex_collector.py

3. Collect KOSHA guidance and optionally PDFs: kosha_collector.py

4. Collect regulated-substance lists: substance_collector.py

5. Normalize and import into SQLite: importer.py

6. Audit generated data: verify_db.py

These are independent Python CLI stages, run manually on an internet-connected machine.

Source Acquisition Boundary

Collectors isolate external-source concerns from database construction:

• collector.py obtains legislation and notices from law.go.kr without an API key.

• annex_collector.py handles separately published annex content, including HWP extraction.

• kosha_collector.py consumes the data.go.kr KOSHA Guide service; a local kosha_config.json API key enables PDF retrieval.

• substance_collector.py imports chemical regulatory data from ICIS.

Optional binary parsing dependencies (pypdf, gethwp, olefile) affect annex/PDF completeness but not the basic legal-text pipeline.

Database Construction and Quality Gate

importer.py is the central integration boundary: it converts heterogeneous collector outputs into data/laws.db, performs legal-text parsing, extracts CAS identifiers, cleans problematic NULL characters, and chunks annex content for retrieval. verify_db.py is the required quality gate, comparing database content with source collection output for loss, omissions, and search spot checks—particularly important because HWP/PDF extraction can truncate content.

Search Distribution Runtime

dist/she-law-db/ is a self-contained consumer workspace rather than a build artifact to execute from the repository root. Its SQLite search CLI, dist/she-law-db/search_law.py, uses the generated database copied to dist/she-law-db/data/laws.db. It supports text, substance/CAS, and law-name lookup and is designed to run without application servers or LLM infrastructure.

Claude Code Integration Boundary

Claude Code behavior is constrained in the distributable workspace:

• dist/she-law-db/CLAUDE.md defines evidence-first response rules.

• dist/she-law-db/.claude/skills/she-law/SKILL.md defines the search skill.

• dist/she-law-db/search_law.py is the retrieval mechanism Claude invokes.

The integration requires cited source provisions; when retrieval yields no evidence, the intended behavior is to report that verification is unavailable rather than synthesize an answer.