faitholopade/aiact-openrag

EU AI Act — structural RAG corpus

A single-file, pre-embedded SQLite corpus of the EU AI Act — Regulation (EU) 2024/1689 — chunked on legal structure only: one chunk per article paragraph, per recital, per annex point, per Article 3 definition. Chapter, section, and article metadata live in columns, not smeared into the text. The artifact is the corpus, not a pipeline — a downloadable database file you query locally, not a hosted service and not an authoritative legal interpretation.

Release

Release	v1.0.0 (prepared 2026-07-16)
Primary artifact	aiact_openrag.db (SQLite, embeddings included)
Schema version	2 (recorded in the meta table)
Chunks	933
Source text	EUR-Lex consolidated version, CELEX 02024R1689-20240712
Source commit	f94a363
Label audit commit	03e7aff (docs/label-audit.md)

Licensing is mixed — no single licence applies indiscriminately to every file. The EU legal text is reused under Commission Decision 2011/833/EU and is not relicensed by this dataset; the original corpus structure, metadata, embeddings and documentation are CC BY 4.0; adapted benchmark data is CC BY 4.0; software associated with the source repository is Apache-2.0. See LICENSE for the path-level division (LICENSES.md is the same text under the repository's canonical filename) and NOTICE for the attribution notice.

This is not legal advice. This corpus is a retrieval artifact for research and engineering. The chunking, tier labels, and metadata are operator-derived with documented rules — they are not authoritative interpretations. In particular, risk_tier is a labelling convention derived from the Act's structure (see docs/derivation.md), not a legal determination. Consult the Official Journal text and qualified counsel for compliance decisions.

Contents

• 933 chunks: 180 recitals, 522 article-paragraph chunks, 68 Article 3 definitions, 163 annex points
• BGE-M3 dense embeddings (1024-dim float32, L2-normalized) for every chunk
• risk_tier / obligation_on labels derived only where the text is unambiguous — every rule cites its provision in docs/derivation.md; NULL rather than guess
• ELI deep links to the exact provision on EUR-Lex
• Staged application dates per Article 113

Source: EUR-Lex consolidated version 02024R1689-20240712 (English). The English consolidated text is character-identical to the OJ text (32024R1689); article numbering is identical in both, and both CELEX ids are recorded. The build verifies, mechanically and fatally: every article 1–113 appears exactly once; every text node of the normative text is consumed by exactly one chunk and survives verbatim into the output (zero silent drops); no empty chunks; Annex III point counts match an independent extraction.

Schema — table EmbeddingContent

column	type	notes
chunk_id	INTEGER PRIMARY KEY
celex_id	TEXT	02024R1689-20240712
content_type	TEXT	recital | article | annex | definition
chapter	TEXT	e.g. CHAPTER III — HIGH-RISK AI SYSTEMS
section	TEXT	e.g. SECTION 2 — Requirements for high-risk AI systems
article_num	TEXT	6, 6(2), 3(34), recital 44, Annex III.4(a)
heading	TEXT	e.g. Article 6 — Classification rules for high-risk AI systems
heading_level	INTEGER	1 recital, 3 article/annex intro, 4 paragraph/point, 5 sub-point
chunk_text	TEXT	the provision text, nothing else
risk_tier	TEXT	direct textual classification only: prohibited (Art 5(1)) | high (Arts 6(1)/6(2) + Annex III, rebuttable per Art 6(3)) | NULL. limited/minimal are reserved, never assigned — operator-derived, see disclaimer
regime_bucket	TEXT	regime association: prohibited-practices | high-risk | transparency | gpai | voluntary-codes | NULL
obligation_on	TEXT	chunk-level operative subject: provider | deployer | importer | distributor | NULL
eli_url	TEXT	deep link to the exact provision
date_in_force	TEXT	EARLIEST date the provision applies, per Art 113 staging (Annex I: NULL) — an application date, not the Regulation's entry-into-force date (1 August 2024)
embedding	BLOB	1024 × float32, raw bytes, L2-normalized (BGE-M3)
continued	INTEGER	0 normally; 1,2,… for parts of a paragraph split at ~1000 tokens

A meta table records the source URL, both CELEX ids, embedding model, source-text licence notice, schema_version = 2, and a *_semantics note for each of the four derived columns (risk_tier_semantics, regime_bucket_semantics, obligation_on_semantics, date_in_force_semantics) stating exactly what each column does and does not claim.

The four derived columns answer four different questions:

• risk_tier — does this chunk's own operative text classify? Narrow and deliberately sparse: only direct textual classifications ("shall be prohibited", "shall be considered to be high-risk").
• regime_bucket — which regulatory regime does this provision belong to? Structural association (chapter/annex membership and the Commission's conventional tier names), independent of whether the text classifies.
• obligation_on — which enum role is the chunk's operative obligated subject? Judged per chunk from its own sentences; NULL for authority-, Commission-, or AI-Office-bound and descriptive chunks.
• date_in_force — earliest application date per Article 113; not the entry-into-force date, and NULL where no single date is unambiguous.

Value counts (933 chunks):

column	population
risk_tier	prohibited 2 · high 35 · NULL 896
regime_bucket	prohibited-practices 10 · high-risk 332 · transparency 7 · gpai 51 · voluntary-codes 4 · NULL 529
obligation_on	provider 56 · deployer 17 · importer 7 · distributor 6 · NULL 847

Decoding the embeddings

import sqlite3, numpy as np

con = sqlite3.connect("aiact_openrag.db")
con.row_factory = sqlite3.Row
row = con.execute("SELECT * FROM EmbeddingContent WHERE article_num = '5(1)'").fetchone()
vec = np.frombuffer(row["embedding"], dtype=np.float32)   # shape (1024,)

Getting started — semantic search in ~15 lines

Encoding new queries needs the embedding model (pip install sentence-transformers); everything else is the standard library plus NumPy.

import sqlite3, numpy as np
from sentence_transformers import SentenceTransformer

con = sqlite3.connect("aiact_openrag.db")
rows = con.execute("SELECT chunk_id, article_num, heading, chunk_text, embedding "
                   "FROM EmbeddingContent WHERE embedding IS NOT NULL").fetchall()
mat = np.vstack([np.frombuffer(r[4], dtype=np.float32) for r in rows])

model = SentenceTransformer("BAAI/bge-m3")
q = model.encode(["is emotion recognition at work prohibited?"],
                 normalize_embeddings=True)[0]
for i in np.argsort(-(mat @ q))[:5]:
    r = rows[i]
    print(f"{r[1]:<16} {r[2]}\n   {r[3][:150]}…\n")

Filter by metadata instead of (or as well as) similarity — that is the point of structural chunking:

-- what the text itself classifies (direct, defensible per provision)
SELECT article_num, chunk_text FROM EmbeddingContent
WHERE risk_tier = 'prohibited';

-- everything belonging to a regime (the broader, structural grouping)
SELECT article_num, heading FROM EmbeddingContent
WHERE regime_bucket = 'high-risk' AND content_type = 'article';

Evaluation

The corpus was evaluated against the AI Act Evaluation Benchmark (Davvetas et al.) on retrieval, QA and risk-classification tasks; full numbers, methodology and provenance are in eval/RESULTS.md.

One finding deserves a plain-language caveat here. A three-family model panel (mistral-nemo, qwen2.5:14b, claude-fable-5) rating all 339 benchmark scenarios showed inter-model agreement collapsing exactly on the two tiers where classification F1 is lowest: per-tier Fleiss κ was 0.238 (limited) and 0.449 (minimal) versus 0.784 (prohibited) and 0.738 (high-risk). This is consistent with ambiguity or noise at the limited/minimal label boundary, and it means limited/minimal per-tier scores on this benchmark should not be read as pure retrieval-quality signals — but it is an exploratory finding with substantial caveats, not proof: the raters are themselves LLMs, so shared model difficulty on a genuinely hard boundary cannot be separated from label noise, and the benchmark's labels are LLM-generation targets under human-authored tier prompts rather than independently validated classifications.

Known limitations

Plain language, so nobody trips on these later:

• Frozen in time. This is the consolidated version of the Act as of its consolidation date (CELEX 02024R1689-20240712), as fetched from EUR-Lex at build time (July 2026). If the Act is amended after that, this corpus does not know. Check EUR-Lex for anything load-bearing.
• Harmonised standards are not in here — on purpose. The CEN/CENELEC standards that operationalise the Act are copyrighted and cannot be redistributed. Commission guidance without an explicit reuse notice is also excluded. A RAG system built on this corpus alone cannot answer "which standard do I follow?"
• risk_tier values limited and minimal never occur — on purpose. The Regulation's operative text never classifies a system into those tiers; they exist only in the Commission's four-tier presentation. An empty result for risk_tier = 'minimal' is correct behaviour, not missing data. The conventional associations are preserved in regime_bucket (transparency ≈ "limited risk", Art 50; voluntary-codes ≈ "minimal risk", Art 95).
• risk_tier is deliberately sparse. Only 37 of 933 chunks carry it (2 prohibited, 35 high) because almost none of the Act's text directly classifies anything — most of it is regime machinery. Use regime_bucket (404 non-NULL) for "provisions of the high-risk/prohibited/GPAI regimes".
• All derived columns are rule-derived, not court-decided. Every rule and the provision it rests on is written out in docs/derivation.md, and the audit that produced the current design in docs/label-audit.md. Where the text is ambiguous the value is NULL. Read the rules before trusting the labels.

Attribution

• Source text: © European Union, 1998–2026, via EUR-Lex. Reuse permitted under Commission Decision 2011/833/EU on the reuse of Commission documents; reproduction with acknowledgement of the source. Only the versions of EU law published in the Official Journal of the European Union are authentic.
• Evaluation benchmark: scenarios and QA pairs from the AI Act Evaluation Benchmark (CC-BY-4.0 data, Apache-2.0 code).

Citation

If you evaluate against the benchmark used in this corpus's evaluation, cite:

@article{Davvetas2026,
  title   = {AI Act Evaluation Benchmark: An Open, Transparent, and
             Reproducible Evaluation Dataset for NLP and RAG Systems},
  author  = {Athanasios Davvetas and Michael Papademas and Xenia Ziouvelou
             and Vangelis Karkaletsis},
  journal = {arXiv preprint arXiv:2603.09435},
  year    = {2026}
}