| name | tw-legal-rag |
| description | Retrieve real Taiwan court judgments with verifiable citations before answering any question about Taiwan law or case law. Use when the user asks about 台灣/臺灣 judgments, 判決, 法院見解, legal precedents, or wants case support for a legal argument. Wraps the twlegalrag CLI (pack subcommand); answers must cite only citation_ids present in the returned bundle. |
TW Legal RAG — Taiwan judgment retrieval for Claude Code
When to use
Any time the question involves Taiwan law and the answer would benefit from
actual court judgments: legal questions in plain language, case-law research,
checking whether a cited 字號 exists, or building an argument that needs
precedent support. Retrieve FIRST, answer second.
How to retrieve
Run the bundled script with the user's question:
python3 scripts/search_judgments.py "雇主未足額提撥勞工退休準備金的法律責任" -n 5
It prints a bundle JSON containing judgments (court, case number, date,
reasoning excerpts), stable citation ids (J1, J2, ...), an
allowed_citations whitelist and verification instructions.
The script locates the twlegalrag executable automatically and falls back
to python -m twlegalrag when the console script is not on PATH (common on
Windows user-site installs). If both fail, install the CLI first:
pip install twlegalrag.
Citation rules (non-negotiable)
- Cite ONLY ids listed in the bundle's
allowed_citations. Judgments in
unread_candidates were not read in full — never present them as
authority.
- Quote only text that appears verbatim in a judgment's
fulltext_excerpt. Do not reconstruct holdings from memory.
- If the bundle is empty or retrieval fails, say so explicitly. Never
substitute invented case numbers.
- A case number appearing in the bundle proves the judgment exists — it
does not prove it supports the user's position. Read the excerpt before
characterising a holding.
- Follow the bundle's own
verification_instructions field; it is part of
the data contract.
Notes
- Retrieval hits the public TLR endpoint (
tlr.dr-legal.com.tw); no API key
is required and no LLM is called server-side.
- Windows: if output shows garbled Chinese, the terminal is not in UTF-8;
the CLI reconfigures its own streams, but pipe output to a file when in
doubt.