KGCache: Amortized Subgraph Retrieval for KG Reasoning with LLMs

2026-08-08

AI Summary

This work addresses the inefficiency in knowledge graph question answering caused by repeatedly retrieving identical subgraphs. The authors propose KGCache, a caching architecture that stores one-hop neighborhood subgraphs within large language model–driven QA systems, supporting both iterative traversal and one-shot planning reasoning paradigms. They present the first systematic analysis of entity reuse and introduce semantic caching to improve hit rates. Efficient cache management is achieved by integrating LRU, LFU, and a trajectory-aware Oracle policy with semantic context matching. Evaluated on WebQSP, the system achieves up to 1.91× retrieval speedup, 1.06× end-to-end acceleration from semantic caching, and a per-hit latency reduction of up to 3.73×.

Abstract

Large language models can answer knowledge-intensive questions more reliably when they are grounded with knowledge graphs, but systems such as Think-on-Graph and Reasoning-on-Graph repeatedly query the same graph neighborhoods across different questions. In this work, we study this repeated retrieval in Knowledge Graph Question Answering~(KGQA) workloads and propose KGCache, an in-memory cache for one-hop knowledge graph neighborhoods. KGCache is designed to be compatible with both iterative traversal (ToG) and one shot planning (RoG) KGQA paradigms. KGCache is placed between the KGQA engine and the backend serving the KG, so repeated entity requests can be served from cache instead of issuing new KG queries. We evaluate KGCache on WebQSP and CWQ using LRU, LFU, and a trace-aware Oracle policy. Our analysis shows that both datasets contain substantial entity reuse among starting entities and entities reached during traversal. We also explore semantic caching for similar queries, which shows additional hit-rate gains on WebQSP and needs further accuracy testing on CWQ. Entity caching accelerates KG retrieval by up to $1.91\times$, while semantic-context caching achieves up to $1.06\times$ full-system speedup in the evaluated WebQSP configurations, with each hit being up to $3.73\times$ faster.

Problem

• Knowledge Graph Question Answering
• Subgraph Retrieval
• Query Caching
• Entity Reuse
• LLM Reasoning

Innovation

• KGCache
• knowledge graph caching
• KGQA
• semantic caching
• amortized retrieval