Conversational graph retrieval augmented generation for legal land advisory services

Abstract

Accessing accurate and comprehensible legal land information in Uganda remains a major challenge, particularly for individuals in rural areas who may face barriers such as limited awareness of legal procedures or lack of access to qualified legal professionals. The problem lies in effectively retrieving and generating context aware legal guidance from diverse, unstructured, and often scanned sources while ensuring that the information is interpretable, linguistically accessible, and robust against document quality issues. AnAIpowered land advisory system was developed that leverages Graph Retrieval Augmented Generation (Graph RAG) and Computer Vision to provide intelligent, context aware legal support. Graph RAG builds upon Retrieval Augmented Generation (RAG) by introducing a structured knowledge graph composed of entities, legal clauses, and inter document relationships, providing more interpretable responses across multiple legal sources, making it especially suitable for complex legal queries that require structured understanding. Computer Vision techniques were applied exclusively to scanned land title documents using deep learning models including CNN, ResNet, VGG16, VGG19, and Vision Transformer (ViT). An ensemble learning approach combining these models via majority voting significantly improved classification robustness, particularly in handling noisy scans, occlusions, or overlapping features. Among the individual models, ViT achieved the highest classification accuracy at 98%, while the ensemble achieved an average F1 score of 97% across all test samples. The semantic search component was implemented using SentenceTransformers and FAISS indexing, while generative question answering was handled by a fine tuned T5 model, trained over 50 epochs, achieving a validation accuracy of 98% with a validation loss of 0.45. User interaction is supported via a multilingual web interface, where system responses are automatically translated to Luganda using voice synthesis at the output stage to improve accessibility. The system demonstrates a scalable, intelligent conversational legal decision support tool for delivering legal land information through natural conversation and offers a practical solution to bridge the information gap in land governance, empowering citizens to make informed decisions regarding land rights and documentation.