Researchers propose a new data representation called labeled incidence structures (LIS) for handling text, knowledge graphs, and hypergraphs in transformer models. LIS encodes each endpoint as a combination of content, role, and relation instance, allowing a single transformer to process these data types natively without flattening. This representation enables role- and relation-aware attention and can improve model performance in certain scenarios.

Researchers compared Knowledge Graph-based augmentation and Retrieval Augmented Generation (RAG) for culturally-related question answering. Graph-RAG, which uses a knowledge graph built from Wikipedia articles, showed competitive results to standard RAG and improved the base LLM's accuracy by 72-78%. This has implications for the development of more inclusive and explainable AI models.

Quanta is a Python library that integrates dense vector search, full-text retrieval, and knowledge-graph traversal into a single API for hybrid retrieval-augmented generation pipelines. This library makes two key design commitments: weighted reciprocal rank fusion for combining signals and treating the graph as a candidate expander rather than a relevance scorer.