QASP: Query-Adaptive Robust Vector Search Policy

Imagine you're building a search engine that needs to quickly find relevant documents from a massive database. The challenge is that different queries can have vastly different requirements for recall, meaning some queries might miss important results while others retrieve too many irrelevant ones. Currently, many systems use fixed parameters for search, which can lead to inconsistent performance — this is what's called high recall variance. When you evaluate these systems, you often look at average recall, but that can hide the fact that some queries perform poorly while others do well. This inconsistency can frustrate users who expect reliable results every time they search.

To address this, the authors propose a method called QASP, which stands for Query-Adaptive robust vector Search Policy. Instead of relying on fixed parameters, QASP predicts how well a search will perform for each specific query by estimating a recall progression curve. This means that for any given recall target, the system can adapt its search strategy without needing to run multiple models or retrain. The approach not only reduces the number of times the system needs to access data but also allows for real-time adjustments based on how well the search is performing compared to expectations. Compared to traditional methods, QASP achieves significantly lower recall variance and can maintain high performance even as data scales, making it a more reliable choice for builders looking to implement effective search solutions.

QASP introduces a novel approach to vector search by predicting recall curves, which is a significant advancement over fixed policies.

The experimental results show consistent improvements in recall variance and query satisfaction, though further validation across diverse datasets would strengthen claims.