EcoRank: Budget-Constrained Text Re-ranking Using Large Language Models (2402.10866v2)

Abstract: LLMs have achieved state-of-the-art performance in text re-ranking. This process includes queries and candidate passages in the prompts, utilizing pointwise, listwise, and pairwise prompting strategies. A limitation of these ranking strategies with LLMs is their cost: the process can become expensive due to API charges, which are based on the number of input and output tokens. We study how to maximize the re-ranking performance given a budget, by navigating the vast search spaces of prompt choices, LLM APIs, and budget splits. We propose a suite of budget-constrained methods to perform text re-ranking using a set of LLM APIs. Our most efficient method, called EcoRank, is a two-layered pipeline that jointly optimizes decisions regarding budget allocation across prompt strategies and LLM APIs. Our experimental results on four popular QA and passage reranking datasets show that EcoRank outperforms other budget-aware supervised and unsupervised baselines.

Budget-Constrained Text Re-ranking with EcoRank: Optimizing Performance within Financial Limits

Introduction to EcoRank

In the landscape of LLMs for text re-ranking, the challenge of balancing cost with performance has emerged as a predominant concern. Traditional approaches, while proficient in enhancing text re-ranking outcomes, do not take into account the financial implications of utilizing state-of-the-art LLMs. This has led to the development of EcoRank, a sophisticated, two-layered pipeline designed to navigate through the constraints of budget while aiming to retain or improve the re-ranking performance. Our comprehensive suite of methods highlights a strategic allocation of budget across different LLMs and prompt strategies, ultimately presenting a budget-aware solution in the domain of text re-ranking.

Problem Statement and EcoRank's Solution

The current state of text re-ranking with LLMs presents a significant financial challenge, primarily due to the costs associated with LLM API calls. This becomes particularly problematic when businesses need to process a large volume of queries daily, where even cheaper LLM alternatives may lead to unsustainable expenses. To address this issue, EcoRank introduces a methodical approach to optimize re-ranking performance within a given budget, thereby enabling the cost-effective utilization of LLMs for text re-ranking tasks.

Methodology

EcoRank employs a novel two-layer strategy which begins with the initial ranking of passages using a high-accuracy, albeit more expensive, LLM API. This step effectively filters out irrelevant passages early in the process, allowing for a more focused allocation of the remaining budget. The subsequent layer leverages a cheaper LLM API for further re-ranking through pairwise comparisons, a method that proves to be less costly and thus enables the processing of a larger subset of passages within the budget constraints. This layered approach, combined with a strategic budget split, ensures not only an efficient utilization of financial resources but also maintains a high quality of re-ranking.

Experimental Setup and Findings

Our extensive evaluation across four popular datasets—Natural Questions, Web Questions, TREC DL19, and DL20—demonstrates EcoRank's superior performance over other budget-aware baselines, both supervised and unsupervised. We observed a gain of 14% on Mean Reciprocal Rank (MRR) and Recall@1 (R@1) metrics, which are significant improvements in the context of cost-aware re-ranking. Additionally, our experiments underlined the importance of selecting appropriate LLM APIs and prompt designs tailored to the budget constraints, further showcasing EcoRank's adaptability and effectiveness in a wide range of scenarios.

Theoretical Implications and Future Directions

The introduction of EcoRank presents a significant advancement in the understanding of budget-constrained optimization within the field of text re-ranking. By navigating the intricate balance between cost and effectiveness, EcoRank sets a new precedent for future research in optimizing LLM-based applications under financial constraints. Moreover, the two-layered pipeline strategy adopted by EcoRank opens up new avenues for exploring hierarchical re-ranking frameworks that could further refine the efficiency and accuracy of LLM-based text re-ranking processes.

Concluding Remarks

EcoRank represents a pivotal contribution to the field of text re-ranking with LLMs, addressing the oft-overlooked aspect of budget constraints. By leveraging a methodological approach that encompasses varying prompt designs, LLM API choices, and strategic budget allocation, EcoRank not only demonstrates a significant enhancement in re-ranking performance but also pioneers the path toward a more financially sustainable application of LLMs. Future explorations could delve into the automation of LLM and budget choices within EcoRank, further optimizing its cost-effectiveness and adaptability to diverse datasets and tasks.