QV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving
QV-PIC tries to make visual caching practical for RAG by restoring lost text detail without giving up speed.
The paper says rendered text images can shrink RAG inputs into fewer visual tokens, but ordinary image-based PIC loses more answer quality than text PIC. QV-PIC compiles visual caches offline with the model’s native chat-template prefix, then uses query relevance to spend high-resolution budget where fine textual evidence matters. Across six tasks, the authors report a 21.6-point average F1 gain over vanilla rendered-image PIC and an 83.8% TTFT reduction versus full prefill. ArXiv · AI/CL/LG's note
The paper says rendered text images can shrink RAG inputs into fewer visual tokens, but ordinary image-based PIC loses more answer quality than text PIC. QV-PIC compiles visual caches offline with the model’s native chat-template prefix, then uses query relevance to spend high-resolution budget where fine textual evidence matters. Across six tasks, the authors report a 21.6-point average F1 gain over vanilla rendered-image PIC and an 83.8% TTFT reduction versus full prefill. ArXiv · AI/CL/LG's note
score 4