About the Project
LexA: Artificial Intelligence for Legislative Document Retrieval and Question Answering
Coordinated by NOVA IMS, LexA addresses a structural bottleneck in public administration: the inability of traditional, keyword-based search tools to cope with the volume, complexity and semantic subtlety of legislative texts. Building on recent advances in Natural Language Processing and Large Language Models, the project develops a proof-of-concept platform that (i) retrieves the most relevant passages of law through dense, context-aware embeddings and a two-stage retrieval pipeline (bi-encoder recall followed by cross-encoder re-ranking), and (ii) answers domain-specific questions posed by civil servants by coupling an LLM-based QA module with an automated “LLM-as-a-judge” quality control layer.
The workflow spans five phases: acquisition and preprocessing of the Diário da República corpus; generation and indexing of contextual embeddings; implementation of the retrieval and re-ranking engine; fine-tuning and evaluation of the QA component; and final integration, testing with the National Institute of Administration (INA, I.P.), and refinement of the user interface for operational uptake.
As coordinator, NOVA IMS applies its expertise in data science and AI to:
- Design and fine-tune bi-/cross-encoder architectures for semantic retrieval and re-ranking.
- Build and index a large-scale legislative corpus, ensuring data quality and compliance with FCT requirements.
- Implement and validate an LLM-based QA module supplemented by an automated quality-control “judge” model.
- Conduct real-world testing with the public administration partner and iterate on usability, reliability and bias-mitigation strategies.
Expected impacts include faster and more consistent responses to citizens, improved evidence-based decision-making, and a transferable architecture for other government domains.
