Evaluation
50 hand-written questions about the AI Act, each with the article or annex that answers it. Every change to search or to the agent is measured against this set. Scripts: scripts/eval_retrieval.py and scripts/eval_agent.py.
The agent vs one search
Search methods (no AI model)
hit@k: share of questions with a correct article in the top k results. MRR: average of 1/rank of the first correct result. Recitals are weighted 0.7 because they explain the law but are not the law; this raised hit@1 from 0.44 to 0.66 (tuned on this same set, so treat it as optimistic).