CAG vs RAG: NACE 2008 → NACE 2025 recodification

LLM-based recoding of business activity descriptions from NACE 2008 to NACE 2025, comparing a Retrieval-Augmented Generation (RAG) strategy against a Context-Augmented Generation (CAG) strategy.

Repository structure

.
├── src/
│   ├── 1_build_vector_db.py          # Build Qdrant collection (RAG only)
│   ├── 2_encode_unambiguous.py       # Rule-based recoding for unambiguous cases
│   ├── 3_encode_ambiguous.py         # LLM-based recoding (RAG or CAG)
│   ├── 4_ensemble_predictions.py     # Majority-vote ensemble across multiple LLM runs
│   ├── 5_build_nace2025_sirene4.py   # Build the final NACE 2025 dataset
│   ├── strategies/                   # rag.py, cag.py, base.py
│   ├── vector_db/                    # Qdrant client + NACE notices loader
│   ├── evaluation/                   # Metrics + reports
│   ├── mappings/, constants/, config/, utils/
│   └── analyse_errors.qmd, analyse_metrics.qmd
├── argo-workflows/                   # Argo workflow specs
├── tests/
├── pyproject.toml
└── .env.example

Setup

Requires Python 3.12+ and uv.

uv sync
cp .env.example .env   # then fill in the secrets
uv run pre-commit install

Running 3_encode_ambiguous.py

RAG (retrieves top-k NACE notices from a Qdrant collection before generation):

uv run src/3_encode_ambiguous.py \
  --strategy rag \
  --llm_name qwen3-6-35b-moe \
  --collection_name embeddings_qwen \
  --experiment_name NACE2025_DATASET \
  --top_k 5 \
  --third 1

CAG (the full NACE notice catalog is passed in the prompt, no retrieval):

uv run src/3_encode_ambiguous.py \
  --strategy cag \
  --llm_name qwen3-6-35b-moe \
  --experiment_name NACE2025_DATASET \
  --third 1

Common flags: --sample_size N for a subset, --thinking to enable LLM reasoning mode, --only_annotated true to restrict to annotated cases. See python src/3_encode_ambiguous.py --help for the full list.

License

Apache License 2.0 — see LICENSE.