Singapore Academy of Law (LawNet)

Singapore's legal profession relied on Boolean keyword search and iterative reading to conduct case law and legislative research — a process that required lawyers to manually scan through lengthy judgments and cross-reference multiple sources before arriving at usable answers. Two out of every three court judgments lacked official summaries, forcing researchers to read full-length documents averaging 8,000 words each. For the approximately 10,000 LawNet users — including the majority of Singapore lawyers in private practice — this represented significant time lost to routine research that could not be redirected toward client counsel or strategic analysis.

IMDA and the Singapore Academy of Law (SAL) jointly developed the GPT-Legal Q&A model, a generative AI-powered search engine deployed within LawNet 4.0. The system combines Large Language Model technology with Retrieval-Augmented Generation (RAG), trained on an authoritative corpus of Singapore-specific legal content: court judgments, Singapore Law Reports, legislation, and legal texts. Unlike its predecessor — the first GPT-Legal model, which generated 800-word headnote-style summaries for judgments lacking them — GPT-Legal Q&A is designed for interactive query resolution. It interprets the intent and context behind legal questions across factual, case-specific, and hypothetical archetypes, and returns precisely referenced answers in real time. The initial release is tuned for contract law, with planned expansion to family and criminal law.

The GPT-Legal Q&A model delivers measurable research efficiency gains across LawNet's user base:

• 75%+ of Singapore lawyers in private practice benefit from a materially shorter legal research journey
• ~10,000 LawNet users now have access to context-aware, real-time answers in place of iterative keyword searches
• 15,000+ court judgments were summarized by the earlier GPT-Legal model, condensing 8,000-word documents into 800-word headnote-format summaries

The shift from exact-term matching to intent-aware retrieval removes the need for researchers to manually formulate Boolean queries and cross-read multiple results — redirecting that time toward higher-value legal analysis and client advisory work.

• RAG architecture is essential for domain-specific AI accuracy — grounding LLM responses in authoritative, curated legal corpora reduces hallucination risk and maintains citation reliability.
• Jurisdiction-specific training data is non-negotiable — generic legal AI cannot substitute for a model trained on Singapore statutes, case law, and law reports.
• Phase deployment by practice area — launching with contract law before expanding to family and criminal law allows quality validation before broader rollout.
• Pair generative AI with structured output formats — mirroring the established headnote structure builds practitioner trust and adoption.
• Measure time-to-insight, not just usage — the headline metric here is research journey length, a more meaningful proxy for professional value than raw query volume.