Overview
QASP (Query-Adaptive robust vector Search Policy) addresses the challenge of achieving high recall while minimizing computational costs in vector search. It operates by predicting the full recall progression curve for individual queries through a single upfront supervised regression. This prediction informs a search policy adaptable to any recall target, bypassing the need for iterative model invocations during search or distinct predictors for each target. QASP predicts normalized recall values using scale-invariant features and pre-search inference, which facilitates its generalization across various recall targets, index configurations, and datasets. The system also integrates a lightweight reactive component that adjusts search depth based on discrepancies between predicted and observed progress, without requiring additional inference steps.
Research Context
A primary challenge in vector search involves maintaining consistently high recall while concurrently minimizing computational expenses. Conventional vector search methods often rely on fixed search parameters, which contribute to considerable performance variance across different queries. The practice of evaluating these systems based on average recall can obscure significant per-query performance disparities. The fixed nature of these parameters can lead to an irreducible lower bound on the loss for any fixed policy.
Approach
QASP employs a supervised regression model to predict the entire recall progression curve for each query. This prediction is conducted upfront, using a single inference step per query. The model utilizes scale-invariant features to predict normalized recall values. This design allows QASP to derive a search policy for any specified recall target from a single prediction, rather than requiring multiple model calls or specialized predictors for different targets.
The policy's generalization capability is attributed to its use of scale-invariant features and pre-search inference. This enables QASP to be applied across various recall targets, index configurations, and datasets without requiring retraining for each specific scenario. The system's architecture includes a reactive component that dynamically modifies search depth. This adjustment is based on real-time deviations between the predicted recall progression and the observed search progress, and it operates without necessitating further inference.
Findings
- QASP's loss was found to exceed the irreducible lower bound of any fixed policy by a vanishing margin.
- The data access savings realized by QASP over fixed probing methods were observed to grow exponentially with intrinsic dimensionality.
- The training sample required for QASP was proven to be finite, irrespective of the dataset size and dimensionality.
- Experimentally, QASP achieved significantly lower recall variance and reduced deviation from the target recall compared to existing methods.
- The system demonstrated a higher query satisfaction rate.
- QASP exhibited scalability to large datasets and hierarchical indices without requiring retraining.
- Specifically, QASP achieved 99% recall with an 80% reduction in data access during experimental evaluations.
Why This Matters
The ability of QASP to predict per-query recall progression and adapt search policies contributes to more consistent and reliable vector search performance. By reducing recall variance and computational overhead, it offers a method for optimizing resource utilization in large-scale data retrieval systems. The demonstrated scalability without retraining suggests potential for efficient deployment in dynamic environments.