NJ Office of the Public Defender Launches Closed AI Legal Retrieval Library

The New Jersey Office of the Public Defender (NJOPD), collaborating with researchers, state innovation agencies, IT departments, AWS engineers, and public defenders, has released an AI-enabled resource library tailored for public defense. Aimed at improving information retrieval in high-pressure, time-sensitive legal aid scenarios, the project represents a practical application of AI research in frontline public services.

Tools and Use Cases

The AI-enabled Resource Library allows NJOPD employees to input legal queries and quickly locate relevant passages within a secure, closed database. The author notes that public defenders handle massive caseloads under tight deadlines, relying on shared knowledge scattered across existing briefs, memos, directives, and peer experiences. The tool's significance lies in organizing and surfacing these internal materials much faster.

Technical Approach and Risk Control

Through structured pre-launch interviews, the team identified hallucination as a primary risk, leading them to adopt an "extractive-first" approach. The system prioritizes extracting facts from original materials before offering optional generative summaries. According to the author, these summaries are strictly for supplementing case facts and aiding comprehension of search results, not for replacing original text or serving as legal claims.

Evaluation and Model Adjustments

The author noted that new evaluations revealed most open-source embedding models underperform in this real-world legal context. Consequently, the team pivoted to fine-tuning models with synthetic data to boost operational efficacy. To align evaluations with frontline needs, they collaborated with real appellate-level public defenders to compile a new set of representative queries, emphasizing that existing academic benchmarks do not fully match actual litigation issues. The team intends to continue applying AI research to enhance public services and hopes these resources will be widely used by NJOPD, other offices, and the public.

2026-07-15 ~ 2026-07-15 · 8 related posts