KGCaRe Beats Think-on-Graph and Vanilla RAG by Traversing an LLM-Built Knowledge Graph Iteratively, Re-Entering With Clue Entities When Context Falls Short

arXiv 2608.09779·medium signalXBlueskyLinkedInCopy link

Posted 2026-08-10, KGCaRe is a hybrid retrieval approach for complex conditional QA that builds a knowledge graph from documents via multi-prompt extraction, stores it in a graph database, and simultaneously embeds the same documents for neural retrieval. Its distinguishing mechanic is LLM-guided iterative graph traversal: extract relevant triples, prune irrelevant ones, and if the first traversal doesn't produce satisfactory context, re-traverse using additional clue entities. Path-form triples plus retrieved passages feed custom prompts that generate answers with explanations. On two complex conditional QA datasets it beats Vanilla LLM, Code Prompt, Text Prompt, Think-on-Graph, Vanilla RAG, and HybridContextQA across Mistral, Mixtral, GPT-3.5, and GPT-4o; the pipeline is publicly released.