Retrieval-Augmented Generation (RAG) augments large language models (LLMs) with external knowledge, yet most RAG systems rely solely on vector-based semantic search, limiting their handling of structured or relational information. We propose a hybrid multi-source RAG architecture that unifies graph-based and vector-based retrieval within a single conversational framework. A prompt-driven orchestrator dynamically selects and fuses the appropriate source at runtime, leveraging graph databases for symbolic reasoning and vector stores for semantic similarity. Experimental evaluation with LLM-as-a-Judge scoring shows that the hybrid configuration consistently outperforms single-source baselines, achieving higher relevance, fluency, faithfulness, and interpretability without any additional training overhead.
Di Nicola, G., Iannucci, S., Torlone, R. (2026). Hybrid Multi-Source RAG for Context-Aware Conversational Data Access. In CEUR Workshop Proceedings. CEUR-WS.
Hybrid Multi-Source RAG for Context-Aware Conversational Data Access
Di Nicola G.;Iannucci S.;Torlone R.
2026-01-01
Abstract
Retrieval-Augmented Generation (RAG) augments large language models (LLMs) with external knowledge, yet most RAG systems rely solely on vector-based semantic search, limiting their handling of structured or relational information. We propose a hybrid multi-source RAG architecture that unifies graph-based and vector-based retrieval within a single conversational framework. A prompt-driven orchestrator dynamically selects and fuses the appropriate source at runtime, leveraging graph databases for symbolic reasoning and vector stores for semantic similarity. Experimental evaluation with LLM-as-a-Judge scoring shows that the hybrid configuration consistently outperforms single-source baselines, achieving higher relevance, fluency, faithfulness, and interpretability without any additional training overhead.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
