v1.0.0 — 正式 stable release

首個正式穩定版（Development Status 升至 Production/Stable）。已發佈到 PyPI：pip install -U twlegalrag

重點變更

• bundle.py 引用防幻覺強化：pack 產出的 bundle 加入 CRITICAL: 最高優先指示 + 自我檢查規則，明確禁止下游 AI 從記憶／網路搜尋／其他資料庫補引用，只准引用 allowed_citations 內的判決字號（codex A+B+C review）。
• 新增使用者指南 docs/USER_GUIDE_zh.md：MCP / CLI step-by-step 接法。
• Remote MCP 文件：ChatGPT 與 Claude 皆透過 Remote MCP connector 接同一端點 https://tlr.dr-lawbot.com/mcp（不再走 GPT Action）；docs/mcp-anchor.md 說明 anchor bundle 流程。

定位（不變）

Retrieval-only：連法律偵探 TLR 公開端點檢索 2,200 萬筆台灣判決 → 打包 bundle 交你自己的 AI → bundle 層級引用檢查。不呼叫任何 LLM、不生成法律意見、不為第三方模型輸出背書。

安裝

pip install -U twlegalrag
twlegalrag health

🤖 Generated with Claude Code

v0.1.0 — first public release

First public (beta) release of Taiwan Legal RAG — an open-source CLI for
semantic Taiwan legal judgment retrieval, powered by Legal Detective's
22M-judgment retrieval infrastructure.

pip install twlegalrag
twlegalrag pack "詐欺和債務不履行怎麼區分" -o bundle.json

What it does

• search — list matching judgments (semantic fuzzy search, not keyword-only).
• pack — bundle retrieved judgments (with stable citation ids, excerpts,
  citation URLs, and verification instructions) for use with your own AI tool.
• check — bundle-level citation check of an AI-generated answer.
• health — check the retrieval service.

What it is / is not

• Retrieval-only: it does not call any LLM and does not generate legal advice.
• Answers are produced by your own AI (ChatGPT / Claude / Gemini / local model).
• The built-in citation check is bundle-level, best-effort: it verifies whether
  cited case numbers are in the retrieved bundle and whether verbatim quotes appear
  somewhere in the bundle's text. It does not verify legal reasoning or guarantee
  semantic faithfulness of third-party model outputs.

Notes

• Your query text is sent to the TLR retrieval endpoint and may be logged for
  retrieval-quality analysis; do not submit personal secrets. See the README.
• The judgment database, embeddings, and retrieval logic stay server-side; this
  repo is the open client + a citation check.

MIT licensed. This is a beta / soft launch.