law-assistant

한국어

---

FAISS-based Retrieval-Augmented Generation (RAG) over the Korean law corpus in law/. Supports embeddings and chat with Ollama, OpenAI, and Anthropic.

Quick start

• Choose a preset and copy to .env:
  • cp .env.ollama .env (local models)
  • cp .env.openai .env (OpenAI only)
  • cp .env.anthropic .env (Anthropic chat + local embeddings)
• Ensure Python 3.10+ is installed.
• Install deps: pip install -e . (or pip install -r requirements.txt).

Build the index

• Default uses Ollama embeddings nomic-embed-text:
  • law-assistant index --chunk-size 800 --chunk-overlap 100
• With OpenAI embeddings:
  • EMBEDDINGS_PROVIDER=openai EMBEDDINGS_MODEL=text-embedding-3-small law-assistant index

Query with RAG

• Default chat via Ollama (CHAT_MODEL from .env):
  • law-assistant query "행정절차법의 처분 기준은?"
• OpenAI chat:
  • CHAT_PROVIDER=openai CHAT_MODEL=gpt-5-mini law-assistant query "부칙 적용 기준은?"
• Anthropic chat:
  • CHAT_PROVIDER=anthropic CHAT_MODEL=claude-3-5-sonnet-20240620 law-assistant query "형법 총칙 요건?"

Notes

• Embeddings providers: ollama, openai, anthropic (Anthropic embeddings may require access to /v1/embeddings).
• Ollama must have an embedding-capable model pulled, e.g., ollama pull nomic-embed-text.
• Index files are under data/ (index.faiss, meta.json).

Agent chat loop

• Build the index first (see above).
• Run: python agent.py (auto-builds the index if missing)
• Verbose logs: python agent.py --verbose (shows retrieval/LLM timing and tool calls)
• Commands: :exit, :reset, :topk <int>

Tool calling

• With CHAT_PROVIDER=openai or anthropic, the agent exposes a retrieve(query, top_k) tool the model can call to fetch relevant passages instead of stuffing context into the prompt.
• With other providers (e.g., ollama), the agent falls back to injecting retrieved context into the user message.
• Use --verbose to see per-step timing and tool calls.

Disclaimer

• This project and any agent/script/model responses do not constitute legal advice and must not be relied upon for legal decisions.
• Laws, precedents, and administrative rules may change; the accuracy, completeness, and currency of the repository data (including the law/ submodule) and generated indices are not guaranteed.
• LLMs can hallucinate. Always verify with original texts and official sources.
• No liability is assumed for any damages arising from use; all responsibility lies with the user.
• Consult a qualified attorney for real legal matters.