📌 CAG vs RAG: Centralized Repository for NACE Revision

📑 Table of Contents

• 📌 CAG vs RAG: Centralized Repository for NACE Revision
  • 📑 Table of Contents
  • 📖 Overview
  • 🚀 Getting Started
    • 🛠 Installation
    • 🏗 Pre-commit Setup
    • Cache LLM model
  • 📜 Running the Scripts
    • ✅ 1. Build Vector Database (if you are in the RAG case)
    • 🏷 2. Encode Business Activity Codes
    • 🔬 3. Evaluate Classification Strategies
    • 📊 4. Build the NACE 2025 Dataset
  • 📡 LLM Integration
  • 🏗 Argo Workflows
  • 📄 License
  • TODO

📖 Overview

This repository is dedicated to the revision of the Nomenclature statistique des Activités économiques dans la Communauté Européenne (NACE).

It provides tools for automated classification and evaluation of business activity codes using Large Language Models (LLMs) and vector-based retrieval systems.

🚀 Getting Started

🛠 Installation

Ensure you have Python 3.12+ and uv or pip installed, then install the required dependencies:

uv sync

or

uv pip install -r pyproject.toml

🏗 Pre-commit Setup

Set up linting and formatting checks using pre-commit:

uv run pre-commit autoupdate
uv run pre-commit install

Cache LLM model

Before running the script download the model and put it in cache using the huggingface CLI which is faster than vllm to download the model.

export MODEL_NAME=Qwen/Qwen2.5-0.5B
uv run huggingface-cli download $MODEL_NAME

📜 Running the Scripts

✅ 1. Build Vector Database (if you are in the RAG case)

To create a searchable database of NACE 2025 codes:

uv run src/build_vector_db.py

🏷 2. Encode Business Activity Codes

For unambiguous classification:

uv run src/encode_unambiguous.py

For ambiguous classification using an LLM:

uv run src/encode_ambiguous.py --experiment_name NACE2025_DATASET --llm_name Qwen/Qwen3-0.6B

🔬 3. Evaluate Classification Strategies

⚠️ TO BE UPDATED

Compare different classification models:

uv run src/evaluate_strategies.py

📊 4. Build the NACE 2025 Dataset

Once all unique ambiguous cases have been recoded using the best strategy, you can rebuild the entire dataset with NACE 2025 labels:

uv run src/build_nace2025_sirene4.py

📡 LLM Integration

This repository leverages Large Language Models (LLMs) to assist in classifying business activities. One can also use all open source models available on HuggingFace and compatible with vLLM.

🏗 Argo Workflows

This project supports automated workflows via Argo Workflows. To trigger a workflow, execute:

argo submit argo-workflows/relabel-naf08-to-naf25.yaml

Or use the Argo Workflow UI.

📄 License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

TODO

1. Mieux versionner les prompts qui sont sauvegardé dans s3. Idealement il faudrait versionner via la collection de la base de données + la version du prompt RAG de langfuse
2. améliorer les embeddings (faut il garder les notes exlicatives en entier pour faire la similarity search ?)
3. implementer le reranker (celui de qwen probablement)
4. Inclure des règles métiers dans les prompts
5. Inclure des règles code spécifique dans le cas du CAG. (Si LMNP alors on explique ce qui fait la distinction entre les deux code -- cf le fichier de @Nathan) --> Du coup inclure des variables annexes pour aider à départager certaines fois ?