MixRAGRec: Mixture-of-Experts Knowledge Graph RAG for Multi-Agent LLM Recommendation
arXiv·medium signal
Combines MoE-based KG retrieval at variable granularities with a three-agent cooperative framework: Retrieval Agent routes queries to appropriate KG experts, Knowledge Preference Alignment Agent converts structured knowledge to LLM-friendly text, and Contrastive Learning Recommendation Agent generates final predictions. Trained end-to-end via novel MMAPO objective. Addresses the one-size-fits-all retrieval problem in KG-RAG systems. Code on GitHub.