An intelligent RAG system that answers French labor law questions by combining general law (Code du travail) with collective bargaining agreements (KALI conventions).

Question: "What is my notice period as an IT engineer with 2 years experience?"

Answer: Under the Syntec convention (IDCC 1486), engineers with 2+ years
        get 2 months notice period, which is more favorable than the
        general labor code minimum of 1 month.

        Sources: Convention Syntec Article 4.3, Code du travail L1234-1

Status: Live on AWS Lambda with Lambda Function URL. Try it at https://zupl7dnbwoqpkamvpxikntjmge0ggnos.lambda-url.eu-west-3.on.aws/

Most legal RAG systems chunk documents arbitrarily. This project preserves legal structure:

• Semantic Chunking: Legal articles kept intact, not split mid-paragraph
• Dual-Source Intelligence: Separate Code du travail and KALI collections enable smart routing
• Convention Detection: LLM automatically identifies which collective agreement applies
• Answer Synthesis: GPT-4o-mini generates natural language answers with citations and confidence scores
• Production Architecture: AWS Lambda + Qdrant Cloud (not just a local demo)

25,798 legal chunks | 7 major conventions | BGE-M3 embeddings | ~€0.0001/query

User Query: "Quelle est la période d'essai pour un ingénieur informatique?"
     │
     ▼
┌─────────────────────┐
│  Routing Agent      │  Detects: IT engineer → Syntec convention (IDCC 1486)
│  (GPT-4o-mini)      │  Strategy: Check Syntec first, then general law
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│  Multi-Collection   │  Searches Qdrant collections with IDCC filter
│  Retrieval          │  Returns: Top 10 relevant articles (scored)
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│  Answer Generator   │  Synthesizes natural language answer
│  (GPT-4o-mini)      │  Tracks citations, confidence score
└──────────┬──────────┘
           │
           ▼
    "As an IT engineer, your trial period under the Syntec convention
     is 3 months (extendable to 4 months with agreement)..."

     Confidence: 0.92 | Sources: Convention Syntec Article 2.1.3

See FLOW.md for detailed architecture decisions and implementation phases.

# 1. Clone and install dependencies
git clone https://github.com/psorianom/admin-rag.git
cd admin-rag
make setup           # Installs Poetry deps + starts Qdrant Docker

# 2. Configure environment
cp .env.template .env
# Edit .env: Add your OpenAI API key for routing/answer generation
# Qdrant: Use local (default) or add cloud credentials

# 3. Load pre-processed data (40MB JSONL files - contact maintainer)
#    OR parse from raw XML (10GB - see Data Sources section)
make ingest-only     # Index pre-embedded vectors (~1 minute)

# 4. Run web UI
poetry run python -m src.retrieval.app
# Open http://localhost:5001

Try a query: "Quelle est la période d'essai pour un ingénieur informatique?"

• Python 3.10+, Docker, Poetry
• OpenAI API key (for routing and answer generation, ~€0.0001/query)
• 40MB disk space (JSONL files) or 10GB (full raw data)

Data: Code du travail + 7 KALI conventions → 25,798 legal article chunks Embeddings: BGE-M3 (1024-dim, French-optimized) via vast.ai GPU (~€0.20) Vector Store: Qdrant Cloud (free tier, 523MB indexed) Routing: GPT-4o-mini detects conventions, applies metadata filters Answer Generation: GPT-4o-mini synthesizes responses with citations Deployment: AWS Lambda + Function URL (currently fixing timeout issues)

Total Cost: ~€0.0001 per query (routing + generation)

See FLOW.md for complete development history, architecture diagrams, and design decisions.

Working:

• Data pipeline: 25,798 chunks parsed and indexed
• Intelligent routing: Convention detection from job roles
• Multi-collection retrieval: IDCC metadata filtering
• Answer generation: Natural language responses with citations
• Local web UI: FastHTML interface on port 5001
• Test coverage: 33 passing tests (routing, retrieval, generation)

In Progress:

• AWS Lambda deployment (timeout issues with API Gateway)
• Switching to Lambda Function URL for longer request times

Coming Soon:

• Public web interface
• Multi-turn conversations
• Enhanced citation links to Légifrance

Intelligent Routing

• Auto-detects applicable collective agreement from job role mentions
• Routes queries to appropriate collections (code_travail, kali, or both)
• Applies IDCC metadata filters for convention-specific results

Multi-Source Retrieval

• Searches both general labor law and conventions
• Understands legal hierarchy (convention rules override general law)
• Merges and ranks results by relevance

Answer Synthesis

• Generates natural language French answers from legal text
• Tracks which sources support each claim
• Provides confidence scores (0-1) with visual indicators
• Highlights cited sources in the UI

Production-Grade Architecture

• Semantic chunking preserves legal article structure
• BGE-M3 embeddings optimized for French legal text
• Efficient ONNX quantized model for CPU inference
• Comprehensive test coverage (33 tests)

Data: Haystack 2.x, BGE-M3 embeddings, Qdrant vector store Intelligence: OpenAI GPT-4o-mini (routing + generation) Deployment: AWS Lambda, Terraform, Docker Development: Python 3.10+, Poetry, Make

See FLOW.md for detailed component decisions and implementation phases.

• FLOW.md: Complete development history, architecture diagrams, and technical decisions
• TODO.md: Project roadmap and task breakdown
• terraform/README.md: AWS deployment guide

Raw Data:

• Code du travail (10GB XML)
• KALI Conventions (10GB XML)

Pre-processed JSONL (40MB, contact maintainer for access):

• code_travail_chunks.jsonl (11,644 chunks)
• kali_chunks.jsonl (14,154 chunks)

Makefile Commands:

make all              # Full pipeline: parse raw XML → embed → index
make ingest-only      # Index from pre-processed JSONL (recommended)
make status           # Check pipeline status

Vast.ai GPU Embedding (optional, if regenerating embeddings):

poetry run python scripts/run_vast_ingestion.py  # ~€0.20 for 25,798 chunks

Immediate: Fix Lambda timeout issues, deploy public web interface Short-term: Multi-turn conversations, enhanced citations to Légifrance Medium-term: Hybrid search (BM25 + semantic), evaluation dataset, more KALI conventions Long-term: Fine-tuned embeddings, PDF export with citations, API authentication

Contributions welcome! This is a research project exploring intelligent legal RAG systems.

Ways to help:

• Report bugs or suggest features (open an issue)
• Share evaluation datasets or test cases
• Improve documentation
• Contribute legal domain expertise

Data: Légifrance via data.gouv.fr Inspiration: AgentPublic/legi for BGE-M3 validation Tools: Haystack, Qdrant, vast.ai, OpenAI

License: MIT (to be added) | Last Updated: January 2026