Ask the EU AI Act. Get answers you can check.
An AI agent searches the official text of Regulation (EU) 2024/1689, reads the articles it needs, answers in plain English and cites every claim. Then it checks that each citation points to a passage it actually read.
Answer
Ask a question or pick an example above. The answer appears here with clickable citations.
Agent trace
- Each search and each article the agent reads shows up here, live.
How it works
Plain retrieval-augmented generation (RAG) searches once and hopes the right passage is in the results. ActLens is agentic: the model decides what to search, reads, searches again with new words if needed, and only then answers.
- Official text, split by structureThe regulation is parsed into 800 passages: article paragraphs, single definitions, annex parts and recitals, each with its exact citation label.
- Hybrid searchKeyword search (BM25) and meaning search (static embeddings) run in parallel and are merged with Reciprocal Rank Fusion. A search takes under a millisecond on a CPU.
- Agent loop with toolsDeepSeek calls
search_lawandreadas tools, up to 6 rounds, and writes the answer with a citation after each claim. - Citation checkEvery citation is matched against the passages the agent actually retrieved. Anything it did not read is flagged, which catches invented article numbers.
Compare the search methods
Free and instant: no AI model involved. Keywords catch exact terms, embeddings catch meaning, and the hybrid merges both. Measured results are on the evaluation page.
Use it inside your own AI app (MCP)
ActLens is also a Model Context Protocol server. Add this address to Claude, Cursor, VS Code or any MCP client, and your assistant gets three tools: search_ai_act, read_ai_act and ask_actlens.
{
"mcpServers": {
"actlens": { "type": "http", "url": "https://actlens.mohitshekhawat.com/mcp" }
}
}