Legal Document RAG — MCP Server

Fully local, offline RAG system for legal documents (Indonesian + English).
Built on MCP SDK, ChromaDB, SQLite FTS5, and multilingual-e5-large.

---

Requirements

• Python ≥ 3.11
• Tesseract OCR with Indonesian language pack:

  # Ubuntu / Debian
  sudo apt install tesseract-ocr tesseract-ocr-ind
  
  # macOS
  brew install tesseract
  brew install tesseract-lang  # includes ind
  

---

Setup

# 1. Clone / copy the project
cd legal_rag_mcp

# 2. Create virtual environment
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. (First run) The multilingual-e5-large model (~2.2 GB) will be
#    downloaded automatically to ~/.cache/huggingface/hub
#    All subsequent runs are fully offline.

---

Usage

Drop documents in the watch folder

legal_rag_mcp/raw_legal_docs/
├── kontrak_jual_beli.pdf
├── perjanjian_kerjasama.docx
├── schedule_of_fees.xlsx
└── scanned_addendum.png

Launch the MCP server

# Development mode (with MCP Inspector UI)
mcp dev mcp_server/server.py

# Production mode (stdio, for Claude Desktop or any MCP client)
python mcp_server/server.py

Claude Desktop configuration (claude_desktop_config.json)

{
  "mcpServers": {
    "legal-rag": {
      "command": "python",
      "args": ["/absolute/path/to/legal_rag_mcp/mcp_server/server.py"]
    }
  }
}

---

Tools Exposed

Tool	Purpose
search_clauses	Hybrid semantic + keyword search
find_conflicts	Cross-document conflict detection (two-zone thresholds)
get_document_overview	Document metadata + clause outline
retrieve_context	Read a clause in full surrounding context
list_documents	Inventory of all ingested files

---

Conflict Detection Thresholds

Zone	Cosine Similarity	Interpretation
Near-duplicate	> 0.88	Same clause — skip
HIGH suspicion	0.75 – 0.88	Same topic, different obligations
MODERATE suspicion	0.60 – 0.75	Related topic, may diverge
Unrelated	< 0.60	Different provisions

Thresholds are configurable in config.py.

---

Re-ingestion (Diff-Based)

When a file in raw_legal_docs/ is modified:

1. File-level SHA-256 hash is compared against the stored hash.
2. If unchanged → skip entirely.
3. If changed → re-chunk the document, compute per-chunk hashes.
4. Only changed or new chunks are re-embedded and upserted.
5. Orphaned chunks (present in old version, absent in new) are deleted.
6. Full re-embedding is never triggered by a partial change.