Source: arXiv

LegalCheck combines Retrieval-Augmented Generation (RAG) with a novel Context-Augmented Generation (CAG) pipeline to draft preliminary legal advice letters for municipal citizen objection cases. The system retrieves relevant laws and prior rulings from a curated knowledge base and allows jurists to upload case-specific documents—such as objection letters and enforcement reports—so that outputs are grounded in both external precedent and the new case facts. Jurists may specify the intended dictum (uphold or reject) and add free-text “steering advice”; controlled prompting and a structured multi-stage prompt workflow integrate these inputs into an initial draft.

A human-in-the-loop reviewer then annotates omissions or suggests revisions, after which LegalCheck performs a refinement pass to incorporate the expert’s feedback. The City of Amsterdam deployment rests on a Zaakbibliotheek of 55,000 past cases and addresses a 35% increase in objection workload reported by the municipal legal department. In the Amsterdam pilot, the paper reports near-final advice letters produced in minutes rather than hours, outputs that capture a large majority (often 80%–100%) of essential legal reasoning, improved consistency and factual accuracy, and positive acceptance by legal professionals while preserving human oversight. The study presents an integrated RAG+CAG architecture, empirical evaluation of speed and quality gains, and discussion of governance, privacy, and reproducibility for responsible public-sector AI adoption.