Utility-based Source Attribution in Retrieval-Augmented Generation

Retrieval Augmented Generation (RAG) systems were introduced to address the knowledge limitations of Large Language Models (LLMs) and their tendency to hallucinate. By leveraging relevant external documents, RAG provides a more grounded generation and access to specific knowledge. Although RAG system are widely used, they still face limitations, such as the inclusion of irrelevant documents in the context and the potential for LLMs to misuse the retrieved information. This work explores effective strategies to enhance RAG systems, with a particular focus on utility-based attribution techniques. Specifically, we investigate the feasibility and effectiveness of adapting Shapley-based attribution to identify influential retrieved documents in RAG. We compare Shapley with more computationally tractable approximations and some existing attribution methods for LLM. This work aims to: (1) systematically apply established attribution principles to the RAG document-level setting; (2) quantify how well SHAP approximations can mirror exact attributions while minimizing costly LLM interactions; and (3) evaluate their practical explainability in identifying critical documents, especially under complex inter-document relationships such as redundancy, complementarity, and synergy.