Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RAG-Stack: Co-Optimizing RAG Quality and Performance From the Vector Database Perspective

Wenqi Jiang

TL;DR

RAG serving seeks to balance generation quality and system performance across a vast configuration space spanning algorithmic and system choices. The authors propose RAG-Stack, a three-pillar framework consisting of RAG-IR workload abstraction, RAG-CM performance cost model, and RAG-PE plan explorer, to identify near Pareto-optimal configurations efficiently. Key contributions include formalizing RAG-IR, detailing cost-modeling approaches (analytical, ML-based, profiling), and outlining a plan-exploration strategy, plus discussion of open research opportunities. The framework aims to provide a reusable, extensible foundation for end-to-end co-optimization of RAG deployments across diverse workloads and hardware stacks.

Abstract

Retrieval-augmented generation (RAG) has emerged as one of the most prominent applications of vector databases. By integrating documents retrieved from a database into the prompt of a large language model (LLM), RAG enables more reliable and informative content generation. While there has been extensive research on vector databases, many open research problems remain once they are considered in the wider context of end-to-end RAG pipelines. One practical yet challenging problem is how to jointly optimize both system performance and generation quality in RAG, which is significantly more complex than it appears due to the numerous knobs on both the algorithmic side (spanning models and databases) and the systems side (from software to hardware). In this paper, we present RAG-Stack, a three-pillar blueprint for quality-performance co-optimization in RAG systems. RAG-Stack comprises: (1) RAG-IR, an intermediate representation that serves as an abstraction layer to decouple quality and performance aspects; (2) RAG-CM, a cost model for estimating system performance given an RAG-IR; and (3) RAG-PE, a plan exploration algorithm that searches for high-quality, high-performance RAG configurations. We believe this three-pillar blueprint will become the de facto paradigm for RAG quality-performance co-optimization in the years to come.

RAG-Stack: Co-Optimizing RAG Quality and Performance From the Vector Database Perspective

TL;DR

RAG serving seeks to balance generation quality and system performance across a vast configuration space spanning algorithmic and system choices. The authors propose RAG-Stack, a three-pillar framework consisting of RAG-IR workload abstraction, RAG-CM performance cost model, and RAG-PE plan explorer, to identify near Pareto-optimal configurations efficiently. Key contributions include formalizing RAG-IR, detailing cost-modeling approaches (analytical, ML-based, profiling), and outlining a plan-exploration strategy, plus discussion of open research opportunities. The framework aims to provide a reusable, extensible foundation for end-to-end co-optimization of RAG deployments across diverse workloads and hardware stacks.

Abstract

Retrieval-augmented generation (RAG) has emerged as one of the most prominent applications of vector databases. By integrating documents retrieved from a database into the prompt of a large language model (LLM), RAG enables more reliable and informative content generation. While there has been extensive research on vector databases, many open research problems remain once they are considered in the wider context of end-to-end RAG pipelines. One practical yet challenging problem is how to jointly optimize both system performance and generation quality in RAG, which is significantly more complex than it appears due to the numerous knobs on both the algorithmic side (spanning models and databases) and the systems side (from software to hardware). In this paper, we present RAG-Stack, a three-pillar blueprint for quality-performance co-optimization in RAG systems. RAG-Stack comprises: (1) RAG-IR, an intermediate representation that serves as an abstraction layer to decouple quality and performance aspects; (2) RAG-CM, a cost model for estimating system performance given an RAG-IR; and (3) RAG-PE, a plan exploration algorithm that searches for high-quality, high-performance RAG configurations. We believe this three-pillar blueprint will become the de facto paradigm for RAG quality-performance co-optimization in the years to come.
Paper Structure (10 sections, 3 figures, 3 tables, 1 algorithm)

This paper contains 10 sections, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. RAG Complexity
  3. Algorithm Configuration Space
  4. System Design Space
  5. Goal: Exploring the Pareto Frontier between RAG Quality and Performance
  6. Proposed Solution: RAG-Stack
  7. RAG-IR as the Workload Abstraction
  8. RAG-CM for Performance Prediction
  9. RAG-PE for Configuration Space Navigation
  10. Conclusion and Future Work

Figures (3)

  • Figure 1: A simple RAG pipeline with models and a database.
  • Figure 2: RAG-Stack systematically explores the performance–quality Pareto frontier for RAG serving.
  • Figure 3: RAG-IR is a collection of dataflow graphs. Each node in the graph is either a database or an inference component.