Retrieval-Augmented Generation (RAG) improves factual grounding, but it also lengthens prompts and raises prefill cost. Prefix caching in serving engines such as vLLM reduces this cost only when requests share the same token prefix. In grounded generation, however, adjacent queries may retrieve overlapping evidence in different orders, so set overlap does not become reusable prefix overlap. We present CacheWeaver, a lightweight prompt-layer method for cache-aware evidence ordering. The method keeps a prefix tree over recently served evidence sequences and uses a greedy walk to place the most reusable prefix first, while leaving the serving engine and retrieved evidence set unchanged. Across three vLLM configurations the method lowers median time-to-first-token (TTFT) by about 20 to 33% relative to retrieval-order prefix caching, without hurting answer quality in our QA tests. The greedy policy reaches 97.5% of the median TTFT gain from oracle ordering, which indicates that most reusable prefix locality can be recovered by a simple scheduling layer between retrieval and inference.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) in Knowledge Graph Question Answering (KGQA) by retrieving reasoning paths from external Knowledge Graphs (KGs). However, the practicality of existing methods is often limited by their reliance on iterative retrieval, where the number of required LLM calls increases with retrieval depth, leading to expensive interaction costs. To address this, we propose Plan-guided Schema-level Reasoning (PSR), a training-free and efficient RAG framework for KGQA that requires only three LLM calls. First, PSR generates a retrieval plan from the input question (the first LLM call). It then performs Plan-Regularized Monte Carlo Tree Search to extract a compact set of schema-level paths from the KG, which are subsequently filtered by the LLM (the second LLM call). Finally, PSR grounds the chosen schema-level paths into instance-level reasoning paths to augment the LLM’s answer generation (the third LLM call). Experimental results on three datasets demonstrate the high efficiency and competitive accuracy of PSR.