sdsforge

GUI + CLI tool for bidirectional conversion between Safety Data Sheet (SDS) documents (Word/PDF) and the Japanese Ministry of Health, Labour and Welfare (MHLW) standard JSON format.

Supports Japanese, English, Simplified Chinese, and Traditional Chinese.

日本語 | 中文

Embedding in your Rust project? Use sdsforge-core directly.

Migrating from sdsconv? See docs/migration-from-sdsconv.md. The old sdsconv binary still works during the deprecation window — it forwards to sdsforge with a warning.

Download

→ All releases & changelogs

GUI Mode

Launch the graphical interface by running sdsforge without any arguments:

sdsforge

The GUI window (820×640) opens with five tabs:

Convert tab	Render tab	Extract Text tab

Drag & drop files onto any tab to fill the input field automatically.

Settings are saved to your OS config directory under sdsforge/config.toml (~/Library/Application Support/sdsforge/config.toml on macOS, ~/.config/sdsforge/config.toml on Linux), migrated automatically from the pre-rename sdsconv/config.toml location the first time you launch this version if the new file doesn't exist yet — your saved API key and settings carry over, and the old file is left untouched. The GUI and CLI share the same conversion engine (tasks.rs), so results are identical.

Commands

to-json — Convert PDF/Word → MHLW standard JSON

# Single file (Anthropic Claude, default)
export ANTHROPIC_API_KEY=sk-ant-...
sdsforge to-json --input input.pdf --output output.json

# Specify source language
sdsforge to-json --input sds_en.pdf --output output.json --lang en

# Batch mode — process a whole directory
sdsforge to-json --input-dir ./pdfs/ --output-dir ./json/ --lang ja

# OpenAI GPT (defaults to gpt-4o-mini)
sdsforge to-json --input input.pdf --output output.json \
  --provider openai --api-key $OPENAI_API_KEY

# Google Gemini (defaults to gemini-2.0-flash)
sdsforge to-json --input input.pdf --output output.json \
  --provider gemini --api-key $GEMINI_API_KEY

# Local LLM via Ollama (any OpenAI-compatible endpoint)
sdsforge to-json --input input.pdf --output output.json \
  --provider local --base-url http://localhost:11434/v1 \
  --model llama3.2 --api-key dummy

# From pre-extracted text (skip PDF parsing)
sdsforge to-json --input extracted.txt --output output.json --lang ja

Provider defaults:

render — Convert MHLW standard JSON → Word / HTML / PDF

# Single file (built-in layout)
sdsforge render --input output.json --to docx --output result.docx --lang ja

# Batch mode (built-in layout)
sdsforge render --input-dir ./json/ --output-dir ./docx/ --to docx --lang en

# HTML or PDF instead of DOCX
sdsforge render --input output.json --to html --output result.html --lang ja
sdsforge render --input output.json --to pdf  --output result.pdf  --lang ja

# Fill a Word template with {{Placeholder}} substitution
sdsforge render --input output.json --to docx --output result.docx \
  --template my_template.docx

# Batch mode with template
sdsforge render --input-dir ./json/ --output-dir ./docx/ --to docx \
  --template my_template.docx

to-docx, to-html, and to-pdf still work as deprecated aliases for render --to docx|html|pdf (same implementation, a deprecation warning is printed to stderr) — switch to render at your own pace.

Word template format

Prepare a .docx file with {{FieldName}} placeholders where FieldName is a leaf key from the MHLW JSON schema. The full dot-path is also accepted for disambiguation.

{{TradeNameJP}}          → 製品和名
{{CompanyName}}          → 会社名
{{Phone}}                → 電話番号
{{IssueDate}}            → 発行日
{{Identification.SupplierInformation.CompanyName}}  → フルパス指定

Placeholders can appear anywhere in the document — paragraphs, table cells, headers, and footers. Word sometimes splits typed text across internal runs; the tool automatically merges such splits before substitution.

extract-text — Extract raw text from PDF/DOCX

Extracts the text that the LLM would receive, without making any API call. Useful for inspecting extraction quality or running the LLM step separately.

# Save to file
sdsforge extract-text --input input.pdf --output extracted.txt

# Print to stdout
sdsforge extract-text --input input.pdf

# Then feed back into to-json
sdsforge to-json --input extracted.txt --output output.json --lang ja

validate — Check a JSON file for structural issues

# Human-readable output (exits 0 = OK, 1 = warnings found)
sdsforge validate --input output.json

# JSON array output for CI/scripting
sdsforge validate --input output.json --json

Checks that key sections (Identification, HazardIdentification, ToxicologicalInformation, etc.) are populated. Exits with code 1 if any issues are found.

Language Support

Requirements

• Rust 1.75+
• An LLM API key (for to-json only) — set the provider's environment variable or pass --api-key
  • Anthropic: ANTHROPIC_API_KEY
  • OpenAI: OPENAI_API_KEY
  • Google Gemini: GEMINI_API_KEY
  • Mistral: MISTRAL_API_KEY
  • Groq: GROQ_API_KEY
  • Cohere: COHERE_API_KEY
  • Local LLM (Ollama etc.): use --provider local --base-url <url> (no API key required)
• Input files must be text-based PDF or DOCX
  • Encrypted PDFs are not supported
  • CID font / Shift-JIS encoded PDFs (common in Japanese documents): handled by pdftotext (poppler) fallback
  • Scanned/image-only PDFs: automatically retried via pdftoppm + tesseract OCR (if installed), or via Claude Vision API (when using --provider anthropic)
  • Full 3-tier PDF fallback: pdf-extract -> pdftotext -> OCR/Vision

Rust Library

The conversion engine is available as a standalone library:

[dependencies]
sdsforge-core = "0.4"

Changelog

Completed in 0.3.6 / 0.2.6

• QC r24: 5 new rule-based checks (S1-ZH-NO-EMERGENCY, S7-FLAMMABLE-STORAGE-TEMP, S8-NO-ENG-CONTROLS, S10-NO-INCOMPATIBLE, CROSS-STALE-DATE)
• QC r24: S8-OEL-NO-NUMERIC false-positive fixes — Chinese unit-before-value format, additional "no OEL" exemption phrases
• QC r24: S5-EMPTY threshold 30→15 chars (reduces false positives for brief Chinese firefighting sections)
• Round-trip test: JSONL parsing fix, validator string-array handling; r24 baseline 30/30 success, CRIT=0, HIGH=9, MED=176
• QC r25: fix S2-EXPLOSIVE-NO-GHS01 / S2-ENV-NO-GHS09 false-negatives (substring "01"/"09" in dates/H-codes); new S3-NAME-IS-CAS (HIGH) and S16-REVISION-BEFORE-ISSUE (HIGH)
• Round-trip test r25 baseline: 30/30 success, CRIT=0, HIGH=13, MED=175
• QC r26: S2-FLAMMABLE-NO-GHS02, S2-CORROSIVE-NO-GHS05, S2-ACUTETOX-NO-GHS06 (all MED) — pictogram/H-code consistency for flammable, corrosive, and acute-tox Cat 1–3; S4-H314-NO-REMOVE-CLOTHING (MED) — P361 compliance
• Round-trip test r26 baseline: 30/30 success, CRIT=0, HIGH=14, MED=181
• LLM prompt: Section 1 Use fallback — source phrase captured when Section 1.2 exists but no specific use is listed (e.g. '无相关详细资料')
• LLM prompt: Section 8 OEL "not required" detection — 不要求 / 无需监控 / 不适用 and similar phrases now stored in AdditionalInfo.FullText instead of being silently omitted
• LLM prompt: Section 9 Densities always extracted; VapourPressure added for flammable/volatile products (H224/H225/H226/H330–H332)
• LLM prompt: Section 12 PersistenceDegradability.BiologicalDegradability always populated when the source subsection exists

References

• MHLW — SDS Standard Data Exchange Format (official page) (Japanese)
• SDS Data Exchange Format Developer Manual (PDF) (Japanese)

License

Licensed under either of:

• Apache License, Version 2.0
• MIT License

at your option.