QV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving
A paper titled 'QV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving' was published on arXiv (cs.AI) on 2026-08-12. It proposes QV-PIC, a query-aware dual-resolution PIC reuse framework guided by model-native templates. The method compiles visual caches offline under the model's native chat-template prefix, and online preserves global context with low resolution while restoring fine-grained textual evidence within a high-resolution budget.
The paper addresses inefficiencies in Retrieval-Augmented Generation (RAG) serving caused by redundant prefill computations. Position-Independent Caching (PIC) reuses precomputed KV caches across positions, but text tokens are voluminous. Rendering text as images compresses tokens but degrades quality due to contextual mismatches and loss of textual detail. QV-PIC uses model-native templates to compile visual caches offline, avoiding online recomputation, and employs a dual-resolution strategy to balance global context and fine-grained evidence.
QV-PIC introduces a query-aware dual-resolution PIC reuse framework. Offline, it compiles visual caches under the model's native chat-template prefix, improving PIC quality without online recomputation. Online, it preserves global context with low resolution and restores fine-grained textual evidence within a high-resolution budget by cumulative attention. This approach mitigates representation-specific quality degradation in rendered-image PIC.
Efficient RAG serving is critical for cost-effective deployment of LLM applications. Reducing redundant prefill computations through visual caching could lower inference costs and latency for retrieval-heavy workloads. The technique may be relevant to providers of RAG infrastructure and enterprise AI platforms.
If validated, QV-PIC could reduce serving costs for RAG systems by decreasing KV cache storage and prefill compute, enabling higher throughput and lower latency for document-grounded AI applications.
Potential next signals include empirical benchmarks comparing QV-PIC against text PIC and other visual PIC methods, open-source implementation releases, and adoption by RAG serving frameworks. Further research may explore generalization to non-chat models or multimodal retrieval.