EU AI Act text split into 933 legally structured, pre-embedded chunks
A new dataset packages the full text of the EU AI Act ( (EU) 2024/1689) for and legal text analysis. Rather than slicing the text into fixed-length windows, it splits along the 's actual legal structure: one chunk per article paragraph, one per recital, one per Article 3 definition, and one per annex point, with chapter, section, and provision metadata stored separately.
The result is a single SQLite file containing 933 chunks, each paired with a 1024-dimensional BGE-M3 embedding. It also includes direct EUR-Lex links, Article 113 -date metadata, and deliberately narrow labels, with direct textual kept separate from broader regulatory-regime association and ambiguous cases left as NULL.
Evaluated against the AI Act Benchmark, the structure-based chunking beat a whole-unit baseline on scenario article recall@20 (0.541 vs 0.449) and QA article hit@10 (0.927 vs 0.898). Overall RAG accuracy was roughly the same or slightly lower with structural chunking, suggesting the generator model's own behavior matters more than chunk granularity for that particular task.
Key points
- EU AI Act text split by legal structure (article paragraphs, recitals, definitions, annex points) into 933 chunks
- Each chunk paired with a 1024-dimensional BGE-M3 embedding, all shipped in one SQLite file
- Includes EUR-Lex links and Article 113 -date metadata; ambiguous s left NULL
- Structural chunking beat baseline on article recall@20 and hit@10 retrieval metrics
- Overall RAG accuracy did not improve, pointing to the generator model as the bigger factor