EU AI Act & Regulation Monitor (RAG-Optimized)

EUR-Lex is free. Keeping a clean, deduplicated, RAG-ready, continuously-monitored corpus of EU AI law is not. This Actor does that maintenance for you, so your LegalTech / compliance / AI product ingests structured Markdown instead of wrestling with EUR-Lex HTML, nested layout-tables, and change tracking.

Built for LegalTech developers, compliance teams, and AI startups building retrieval-augmented (RAG) assistants over EU legislation.

---

What you actually get

• Clean Markdown, not an HTML dump. EUR-Lex renders documents with hundreds of nested <table> elements for layout. This Actor flattens those numbered/lettered provisions into readable lists, preserves genuine data tables as Markdown pipe tables, and strips all site chrome. The EU AI Act comes out as ~115k clean cl100k_base tokens.
• Reliable, structural chunk hints. Article, Chapter, Section, and Annex boundaries are extracted from EUR-Lex's own DOM anchors (not regex over prose), each with the article heading ("Article 6 — Classification rules for high-risk AI systems") and an accurate character offset. Split at chunkHints and ingest — no extra parsing.
• Citation-grade metadata. Every record ships celexId, the ELI URL (the stable, citable URI), documentType, OJ reference, ISO publication date, language, and a consolidated-vs-original flag.
• Token counts included. estimatedTokens (cl100k_base, with a tokenCountMethod field so you always know whether it's exact or approximated).
• Real change detection. State persists across runs. Each document carries a versionHash (SHA-256 of the body), and scheduled runs deliver — and bill — only documents that are new or have changed. You are never re-charged for unchanged law.

Why not just use the free EUR-Lex API?

You can. Here's what you'd be building and maintaining yourself:

	This Actor	Free EUR-Lex API + your code
Discovery across multiple AI search terms	✅ built-in, deduplicated	You build it
HTML → clean Markdown	✅	You build it
Nested layout-tables → readable lists	✅	You build it (this is the hard part)
Real data-tables → Markdown tables	✅	You build it
Article-level chunk hints + headings	✅	You build it
Token counts (cl100k_base)	✅	You build it
Change detection across runs	✅ persistent	You build state management
Ongoing maintenance as EUR-Lex changes	✅ we maintain it	You own it forever

If your team would rather spend an afternoon wiring up a feed than a week building and maintaining a legal-text pipeline, that's the trade.

---

💰 Pricing

Pay-per-event. You are billed per document delivered to the dataset — see the current per-result price on the Apify Store listing. What makes it cheap to run continuously:

• Documents skipped by the relevance filter → free
• Duplicate CELEX IDs across search terms → free (deduplicated)
• Unchanged documents on a scheduled run → free (changedOnly, on by default)

A first full sweep of the AI-relevant corpus is typically a few dozen documents. After that, scheduled monitoring only bills when legislation actually changes — usually a handful of documents a month — instead of re-charging you for the whole corpus every run.

Try it cheaply: the default input (1 term, ~5 documents) lets you see the exact Markdown and chunk-hint output before scaling up.

---

⚙️ Configuration

Parameter	Default	Description
searchTerms	["artificial intelligence"]	EUR-Lex queries. Add more for a broader sweep; duplicates are saved once.
searchText	"artificial intelligence"	Single-term fallback when searchTerms is empty.
searchIn	Title and full text	Search scope.
language	en	Two-letter EUR-Lex language code (structure is identical across all 24).
maxPages	3	Result pages per search term (~10 results each).
maxDocuments	25	Hard cap on documents saved (and billed). The crawl stops exactly here.
changedOnly	true	Deliver/bill only new or changed documents since the last run.
excludeCorrigenda	true	Filter out correction notices.
includeConsolidated	false	Include consolidated texts (no legal effect; documentation only).
maxConcurrency	5	Concurrent HTTP requests.
useApifyProxy	false	Route via Apify Proxy if EUR-Lex ever rate-limits the run's IP.

---

🛠️ Output Schema

{
  "url": "https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689",
  "eliUrl": "https://data.europa.eu/eli/reg/2024/1689/oj",
  "celexId": "32024R1689",
  "title": "REGULATION (EU) 2024/1689 ... (Artificial Intelligence Act)",
  "documentType": "Regulation",
  "ojReference": "OJ 2024/1689",
  "publicationDate": "2024-07-12",
  "language": "en",
  "isConsolidated": false,
  "markdown": "REGULATION (EU) 2024/1689...\n\n## Article 6 — Classification rules for high-risk AI systems\n\n1.   Irrespective of...",
  "estimatedTokens": 115528,
  "tokenCountMethod": "cl100k_base",
  "versionHash": "f4fb8c548d4799f4",
  "scrapedAt": "2026-06-27T16:21:00.000Z",
  "metadata": {
    "chunkHints": [
      { "type": "chapter", "title": "CHAPTER III — HIGH-RISK AI SYSTEMS", "index": 343584 },
      { "type": "article", "title": "Article 6 — Classification rules for high-risk AI systems", "index": 348901 }
    ],
    "totalArticles": 113,
    "totalSections": 162,
    "wordCount": 90698,
    "language": "en",
    "suggestedSplitStrategy": "Split at each Article heading (chunkHints where type=\"article\")."
  }
}

---

🔁 Recommended Usage Pattern

1. Initial sweep — run with your search terms and a maxDocuments budget to build a baseline corpus. (changedOnly: false if you want the whole corpus on the first run.)
2. Daily monitoring — schedule a run with changedOnly: true. It re-checks the corpus and delivers/bills only what changed.
3. RAG ingestion — split each document at chunkHints (type article), carry the enclosing chapter heading as context, and upsert with celexId + eliUrl + publicationDate as metadata filters.

---

What this Actor does not do (yet)

• It defaults to a single language per run (rerun with a different language for parallel versions — anchors align across languages).
• It ships the original adopted act by default. Set includeConsolidated: true for in-force consolidated versions; they are flagged with isConsolidated.
• Relevance filtering is heuristic (title match, or topical density across your search terms + an AI-vocabulary list). Tune searchTerms and maxDocuments to control precision and spend.

---

🏗️ Tech

Plain HTTP + Cheerio (no headless browser). EUR-Lex serves its documents and search results to a normal HTTP client, so a typical run finishes in seconds, not minutes — and there's no brittle bot-detection layer to fight.