Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation

Chao Jin, Zili Zhang, Xuanlin Jiang, Fangyue Liu, Xin Liu, Xuanzhe Liu, Xin Jin

TL;DR

This paper tackles the latency and compute challenges of retrieval-augmented generation by introducing RAGCache, a multilevel dynamic caching system that stores and shares intermediate key-value states of retrieved documents. A knowledge tree organizes KV tensors across GPU and host memory, guided by a prefix-aware PGDSF replacement policy and complemented by cache-aware reordering and dynamic speculative pipelining to overlap retrieval and inference. The authors implement a RAGCache prototype on top of vLLM and Faiss, and show up to 4x faster time-to-first-token and up to 2.1x higher throughput compared with strong baselines, with additional gains over state-of-the-art LLM KV-cache approaches. The results demonstrate significant practical improvements for scalable RAG deployments, especially under skewed retrieval patterns and high-demand workloads.

Abstract

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation

TL;DR

This paper tackles the latency and compute challenges of retrieval-augmented generation by introducing RAGCache, a multilevel dynamic caching system that stores and shares intermediate key-value states of retrieved documents. A knowledge tree organizes KV tensors across GPU and host memory, guided by a prefix-aware PGDSF replacement policy and complemented by cache-aware reordering and dynamic speculative pipelining to overlap retrieval and inference. The authors implement a RAGCache prototype on top of vLLM and Faiss, and show up to 4x faster time-to-first-token and up to 2.1x higher throughput compared with strong baselines, with additional gains over state-of-the-art LLM KV-cache approaches. The results demonstrate significant practical improvements for scalable RAG deployments, especially under skewed retrieval patterns and high-demand workloads.

Abstract

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.
Paper Structure (19 sections, 1 theorem, 5 equations, 19 figures, 4 tables, 2 algorithms)

This paper contains 19 sections, 1 theorem, 5 equations, 19 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. RAG System Characterization
  4. Performance Bottleneck
  5. Optimizations Opportunities
  6. RAGCache Overview
  7. RAGCache Design
  8. Cache Structure and Replacement Policy
  9. Cache-aware Reordering
  10. Dynamic Speculative Pipelining
  11. Implementation
  12. Evaluation
  13. Overall Performance
  14. Case Study
  15. Ablation Study
  16. ...and 4 more sections

Key Result

Theorem 5.1

Let $d$ be the speculative time interval and $T$ be the serving time distribution for the LLM in the G/G/1 request pool, satisfying $T \geq d$ for all serving events. The optimal strategy is: start a speculative generation if the request pool is empty after a stage ends with a different document seq

Figures (19)

  • Figure 1: RAG workflow.
  • Figure 2: Inference time with different input lengths.
  • Figure 3: The distribution of token number.
  • Figure 4: Prefill latency characterization.
  • Figure 5: Retrieval pattern on different datasets.
  • ...and 14 more figures

Theorems & Definitions (1)

  • Theorem 5.1