aa0101181514/tw-legal-rag

What it solves

Taiwan Legal RAG provides a way to retrieve Taiwan court judgments and administrative rules using natural language, solving the problem of traditional keyword-only search tools that often miss relevant cases based on conceptual similarity. It specifically targets the "hallucination" problem in legal AI by providing a structured "bundle" of verified documents that AI models can be forced to cite from, along with a tool to verify if the AI's answer actually uses the documents provided.

How it works

The project consists of a CLI tool and an MCP (Model Context Protocol) server that connect to a hosted semantic retrieval backend containing over 22 million Taiwan court judgments.

1. Retrieval: The tool uses semantic search to find cases that are conceptually similar to a query, even if the wording differs.
2. Packaging: The pack command creates a bundle.json file containing the query, judgment excerpts, and strict verification instructions. This bundle is then fed into an AI (like Claude or ChatGPT) as context.
3. Verification: The check command performs a deterministic string analysis on the AI's output to ensure that any cited case numbers exist within the original bundle and that the quoted text actually appears in the bundle.

Who it’s for

Legal professionals, researchers, and AI developers who need to ground AI-generated legal analysis of Taiwan law in actual, verifiable court judgments without needing to manage their own vector database or LLM infrastructure.

Highlights

• Massive Dataset: Access to 22.5 million+ Taiwan court judgments and thousands of administrative rules.
• Semantic Search: Goes beyond keywords to find conceptually related cases.
• Case History: Includes case_history flags to show if a judgment has been overturned by a higher court.
• Hallucination Guardrails: Uses a read-whitelist and a bundle-level citation check to prevent AI from inventing fake case numbers or legal opinions.
• Zero-Config Integration: Supports Remote MCP for direct integration into Claude and ChatGPT, or a CLI for custom workflows.