From Hours to Minutes: AI for Legal Research
Consejo de la Judicatura · Defensoría Pública del Ecuador · Quito · October 23, 2025
October 23, 2025 was a double day: morning at the Consejo de la Judicatura and afternoon at the Defensoría Pública del Ecuador, both in Quito, with the same material but different audiences. Judges, public defenders, and technical teams shared the same central question: how can artificial intelligence reduce legal research time from hours to minutes?
The answer lies in RAG (Retrieval Augmented Generation) architecture: a system that combines semantic search with language models to find relevant jurisprudence, applicable precedents, and legal doctrine in seconds — not hours. The presentation slides are available so you can view them and follow along at your own pace: view presentation.
The talk covered three axes. First, the diagnosis: thousands of annual rulings, scattered repositories, and searches that consume 4–8 hours per case. Second, the technical architecture: how semantic embeddings represent the meaning of legal texts, how cosine similarity identifies relevant precedents, and why embedding models provide a strong foundation for semantic understanding applied to Ecuadorian legal language. Third, concrete use cases: similar jurisprudence search, applicable precedent identification, detection of inconsistencies between courts, and democratization of citizen access to legal information — available 24/7, in natural language, at no cost.
Prompt Engineering: the interface between human and AI
A significant portion of the session addressed prompt engineering — the discipline of designing precise instructions to get useful results from language models. Having a good model is not enough: how you frame the question largely determines the quality of the answer.
Attendees were given access to the full guide in PDF format and the public prompts repository with a curated catalog of instructions for professional use cases:
→ Prompt Engineering Guide (PDF)
Post-session technical deep-dive: real implementation problems
The most valuable part of the Consejo de la Judicatura visit happened after the formal conference. The institution's technical team requested an extended session to dig into the concrete implementation challenges they face day to day.
The conversation was candid and technical: infrastructure challenges for vectorizing large document volumes, chunking strategies for long legal texts, handling PDFs with irregular structure, latency in production queries, and criteria for evaluating result relevance. I shared direct experience with each of these points — not from theory, but from production systems.
These post-session discussions are often where the greatest value is generated: the space where real questions emerge and where practical experience makes the difference over academic approaches.
Defensoría Pública del Ecuador
The afternoon of the same day, the session at the Defensoría Pública covered the same material with a focus adapted to the public defense context: using AI to research cases with limited resources, the potential to democratize jurisprudence access for defenders in remote areas, and the possibility of conversational systems to guide citizens without legal representation. The talk was broadcast nationwide via Zoom, with more than 50 people connected online from different provinces across Ecuador.
The closing was the same at both institutions: AI does not replace the judge or the defender. It gives them tools to find what they need, faster and more precisely. The time freed up is time to think, analyze, and decide.