Temporal Knowledge Graph-Driven Multi-Agent Collaborative Retrieval Augmented Generation Framework

Retrieval augmented generation (RAG) expands large language models' knowledge by incorporating external evidence, but conventional RAG frameworks often struggle with reasoning over time-stamped events and managing noise from overly broad retrieval. To address these challenges, we propose TKG-MRAG, a temporal knowledge graph-driven multi-agent collaborative RAG framework that tackles both issues. In the indexing phase, LLM extracts 〈entity, relation, timestamp〉 triples from raw text to build a temporal knowledge graph and constructs two-layer indexing structure (the temporal index and the semantic index) to prevent reasoning errors arising from time conflicts. During retrieval, we designed a multi-agent collaborative retrieval mechanism: a temporal planning agent decomposes the query into a sub-query reasoning sequence; a causal retriever agent prunes the temporal subgraph and selects candidate nodes; and a reflective refiner agent ensures consistency, eliminates redundancy, and outputs a compact evidence subgraph. Finally, the original query, sub-query reasoning sequence, and refined subgraph are fed into an LLM to generate the answer. Experiments on the Time-LongQA benchmark demonstrate that TKG-MRAG significantly outperforms baseline methods including Naive-RAG, Light-RAG, and Graph-RAG, achieving superior accuracy in temporal question answering tasks.