Open, structurally chunked EU AI Act corpus released for legal RAG
An open, 933-chunk SQLite dataset splits the EU AI Act by its real legal structure, with embeddings and EUR-Lex links for RAG and legal engineering.
A new open dataset called EU AI Act OpenRAG turns Regulation (EU) 2024/1689 into a downloadable SQLite corpus built for legal research and engineering. Instead of splitting the text into arbitrary token or character windows, as most RAG pipelines do, each chunk follows the Act's actual legal structure: article paragraph, recital, definition, or annex point, with chapter, section, and provision metadata preserved.
The corpus includes 933 chunks, embeddings, exact EUR-Lex links, and metadata on application dates and covered operators. The author took a deliberately conservative approach to legal labeling: a provision is marked as directly classifying a practice or system only when its own operative wording explicitly does so. Broader associations with the prohibited-practices, high-risk, transparency, GPAI, or voluntary-code regimes are stored separately, and ambiguous cases are left as NULL rather than forced into a category.
This distinction matters for engineers building RAG-based compliance tools or legal assistants, since it produces a source of truth grounded in the Regulation's actual legal architecture rather than naive text splitting. Every derivation rule is documented, and the final classification logic was independently reviewed against the Regulation before release. The author is explicit that this is a research and engineering artifact, not legal advice or an automated compliance determination.