aa0101181514/tw-legal-rag

TL;DR

Taiwan Legal RAG is a command‑line tool that lets you query a public, server‑hosted semantic search service containing ~22 million Taiwanese court judgments, statutes and administrative rulings. It returns the matching documents as a bundle that you can feed to any LLM (ChatGPT, Claude, Gemini, a local model, etc.) and it also provides a lightweight, deterministic citation‑check that verifies the LLM’s answer only cites material that is actually present in the bundle.

---

What it is

• Type: Open‑source Python CLI (installable via pip install twlegalrag).
• Purpose: Retrieve Taiwanese legal texts (court judgments, statutes, administrative interpretations) via natural‑language queries and package the results for downstream AI use.
• Scope: Legal‑retrieval + citation verification; does not run or host any language model, nor does it give legal advice.

---

Core features

Feature	What it does	Why it matters
Semantic (fuzzy) search	Uses a server‑side vector‑based index to find judgments that are conceptually related to your query, not just keyword matches.	Lets you discover relevant cases even when you use different wording.
Exact‑case lookup	Precise case‑number lookup with automatic “case not found” handling.	Guarantees you are referencing the correct judgment.
Case‑history linkage	Each judgment bundle includes case_history showing appellate relationships and whether a decision has been overturned.	Prevents citing a judgment that is no longer good law.
Law article lookup	Retrieve the current text of statutes, with amendment dates and validity flags.	Gives you the exact statutory wording to cite.
Administrative‑rule search	Both exact‑number and semantic search for regulations and interpretive orders.	Covers the full hierarchy of Taiwanese law.
Bundle creation (twlegalrag pack)	Packages query, matching judgments, excerpts, URLs, citation whitelist, and verification instructions into a JSON file.	Provides a self‑contained artefact that any LLM can consume safely.
Citation check (twlegalrag check)	Deterministic string‑match check that verifies: (1) cited case numbers are in the bundle, (2) cited text actually appears somewhere in the bundle, (3) no out‑of‑bundle citations.	Helps catch “hallucinated” citations before the answer is presented.
Health endpoint	Simple twlegalrag health ping to confirm the remote service is reachable.	Quick sanity‑check for scripts.
Zero‑LLM dependency	Only requires httpx, typer, rich. No API keys needed for the public endpoint.	Easy to install and run in any environment.

---

How it works (high‑level flow)

1. Search – The CLI sends your natural‑language query to the hosted TLR REST endpoint (https://tlr.dr-legal.com.tw). The server returns a list of matching judgments with a result_token.
2. Full‑text fetch – Using the token, the CLI pulls excerpts (and optionally the full text) for each judgment.
3. Pack – All retrieved data are assembled into bundle.json containing:
   • query
   • citation_id/citation_text/citation_url
   • case_history
   • allowed_citations whitelist
   • verification_instructions (rules for the downstream model to follow).
4. Consume – You paste the bundle into your LLM prompt and ask it to answer using only the provided citations.
5. Check – After the LLM replies, run twlegalrag check bundle.json answer.txt to verify that every citation the model used is present and that quoted snippets exist in the bundle.

---

Who might use it

• Legal tech developers building chat‑bots or RAG pipelines for Taiwanese law.
• Lawyers or researchers who want a reproducible, citation‑safe way to feed case law into LLMs.
• Compliance teams needing to ensure AI‑generated advice references only verified statutes/judgments.
• Educators / students exploring AI‑augmented legal research without handling massive datasets locally.

---

Limitations & caveats

• The CLI is only a client; the heavy lifting (vector embeddings, index, retrieval) lives on the remote TLR server.
• Citation check is string‑based; it cannot confirm that a quoted passage is from the correct judgment or that the legal reasoning is accurate.
• No personal data or confidential documents are ever sent to the TLR service – only the plain query string.
• The service is free and open, but the backend may log queries for quality analysis; do not submit sensitive information.
• The repository is under Elastic License 2.0 (commercial‑friendly) and does not accept external pull requests.

---

Quick start example

# Install
pip install twlegalrag

# Find recent cases about overtime pay
twlegalrag search "勞資 加班費" -n 5 --read

# Create a bundle for an AI prompt
twlegalrag pack "車禍對方全責,我可以求償什麼?" -o bundle.json

# After getting an answer from your LLM, verify citations
twlegalrag check bundle.json answer.txt

---

License & governance

• CLI code: Elastic License 2.0 (free for commercial and internal use, with two restrictions – no re‑hosting the software as a service and must retain the license notice).
• Data & API: Not covered by the code license; usage governed by the separate TERMS.md and TRADEMARK.md files.
• The project is maintained by a single author (Aaron Huang, founder of Legal Detective / Dr.Legal) and does not accept outside pull requests.

---

Bottom line

twlegalrag gives you a practical bridge between a massive, up‑to‑date Taiwanese legal corpus and any LLM you choose, while providing a deterministic, open‑source guard against fabricated citations. It’s a genuine, production‑grade tool for legal‑RAG workflows, not just a demo or a curated list.