LegalBot RAG

Sistema RAG para derecho inmobiliario español con citas obligatorias y validación de fuentes.

🚀 Demo en Vivo

URL: https://legalbot-rag-production.up.railway.app

# Probar el endpoint de chat
curl -X POST https://legalbot-rag-production.up.railway.app/chat \
  -H "Content-Type: application/json" \
  -d '{"question": "¿Cuánto preaviso necesito para no renovar un contrato de alquiler?"}'

También disponible la colección de Postman para pruebas interactivas.

---

Overview

LegalBot is a Retrieval-Augmented Generation (RAG) system that answers questions about Spanish real estate law. It uses:

• Semantic chunking by article (not fixed tokens) to preserve legal structure
• Vector embeddings for semantic search (OpenAI text-embedding-3-small)
• GPT-4 for response generation with mandatory citations
• Response validation to verify citations exist in retrieved context

Architecture

┌─────────────────────────────────────────────────────────────┐
│ INGESTION (offline)                                         │
│ Legal Docs → Chunker (by article) → Embeddings → Chroma     │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│ QUERY (runtime)                                             │
│ Question → Embedding → Similarity Search → Rerank           │
│ → Prompt with context → GPT-4 → Validated Response          │
└─────────────────────────────────────────────────────────────┘

Quick Start

Prerequisites

• Node.js 20+
• pnpm
• OpenAI API key

Setup (Local Vector Store - Sin Docker)

# Clone and install
git clone https://github.com/GMNAPI/legalbot-rag.git
cd legalbot-rag
pnpm install

# Configure environment
cp .env.example .env
# Edit .env and add your OPENAI_API_KEY

# Start the server (usa vector store local pre-cargado)
pnpm dev

Setup Alternativo (Con Chroma)

# Start Chroma (vector database)
docker-compose up -d chroma

# Ingest legal documents
pnpm ingest

# Start the server
pnpm dev

API Usage

# Chat endpoint (local)
curl -X POST http://localhost:3000/chat \
  -H "Content-Type: application/json" \
  -d '{"question": "¿Cuánto preaviso necesito para no renovar un contrato de alquiler?"}'

# Chat endpoint (producción)
curl -X POST https://legalbot-rag-production.up.railway.app/chat \
  -H "Content-Type: application/json" \
  -d '{"question": "¿Cuánto preaviso necesito para no renovar un contrato de alquiler?"}'

# Response includes citations and confidence level
{
  "answer": "Según la LAU, si eres arrendatario debes avisar con al menos 2 meses...",
  "sources": [
    { "article": "Artículo 10", "law": "LAU", "excerpt": "..." }
  ],
  "confidence": "high"
}

Demo Questions

1. "¿Cuánto preaviso necesito para no renovar un contrato de alquiler?"

   • Should cite LAU Art. 10

2. "¿Quién paga las obras de mantenimiento en un piso alquilado?"

   • Should cite LAU Art. 21 and 22

3. "¿Puedo poner un negocio en mi piso de una comunidad de propietarios?"

   • Should cite LPH Art. 7

4. "¿Cuánto es el ITP en Barcelona?"

   • Should respond "No tengo información..." (validates guardrails)

Project Structure

legalbot-rag/
├── src/
│   ├── ingestion/
│   │   ├── chunker.ts        # Semantic chunking by article
│   │   ├── embedder.ts       # OpenAI embeddings
│   │   └── documentLoader.ts # PDF/TXT loading
│   ├── retrieval/
│   │   ├── vectorStore.ts    # Chroma integration
│   │   └── reranker.ts       # Keyword-based reranking
│   ├── generation/
│   │   ├── promptBuilder.ts  # Prompt construction
│   │   └── responseValidator.ts # Citation validation
│   ├── api/
│   │   └── chat.ts           # Fastify endpoints
│   ├── rag.ts                # Main RAG pipeline
│   └── index.ts              # Server entry point
├── data/
│   └── laws/                 # Legal documents (LAU, LPH)
├── tests/
│   ├── chunker.test.ts
│   └── responseValidator.test.ts
├── docker-compose.yml
└── README.md

Key Design Decisions

1. Semantic Chunking by Article

Instead of fixed-size token chunking, we chunk by legal article:

// Bad: Fixed token chunks break legal meaning
"...plazo mínimo de cin" | "co años, o siete si..."

// Good: Semantic chunks preserve full articles
[Artículo 9] "La duración del arrendamiento será libremente pactada..."

2. Mandatory Citations

The system prompt forces citations to reduce hallucinations:

REGLAS ESTRICTAS:
1. SOLO responde basándote en el CONTEXTO LEGAL
2. SIEMPRE cita: [Artículo X, Ley Y]
3. Si no hay info, di "No dispongo de información..."

3. Response Validation

Every response is validated:

• Citations are extracted with regex
• Validated against retrieved chunks
• Confidence level assigned (high/medium/low/none)

4. Overlap Context

Adjacent articles are included as context to improve retrieval for questions spanning multiple articles.

API Endpoints

Method	Endpoint	Description
POST	/chat	Main chat endpoint
POST	/retrieve	Debug: get retrieved chunks
GET	/health	Health check
GET	/stats	System statistics

Configuration

Variable	Default	Description
OPENAI_API_KEY	-	Required
EMBEDDING_MODEL	text-embedding-3-small	Embedding model
CHAT_MODEL	gpt-4-turbo-preview	Chat model
MAX_CHUNKS	5	Max chunks per query
SIMILARITY_THRESHOLD	0.7	Min similarity score
CHROMA_HOST	localhost	Chroma host
CHROMA_PORT	8000	Chroma port

Testing

# Run tests
pnpm test

# Watch mode
pnpm test:watch

# Run evaluation system
pnpm eval

# Detailed evaluation output
pnpm eval --verbose

See docs/evaluation.md for details on the evaluation system.

Production Considerations

For a production system like MIKE, I would add:

1. Observability: OpenTelemetry + Grafana for tracing
2. Feature Flags: A/B test prompts without deployments
3. LLM Evaluation: Golden dataset + automatic quality metrics
4. API Versioning: OpenAPI contracts, auto-generated SDKs
5. Caching: Embeddings cache, response cache for common queries
6. Rate Limiting: Per-client rate limits, queue for heavy queries

License

MIT - Technical demo for interview purposes.