High-dimensional vector retrieval systems often suffer from the curse of dimensionality, leading to reduced efficiency and degraded performance in large-scale applications.
We introduce a goal-aware sparsity framework that learns
adaptive feature masks aligned with specific retrieval objectives,
enabling both computational efficiency through dimensionality
reduction and improved effectiveness by focusing on task-relevant
subspaces. Our approach integrates seamlessly with FAISSbased multi-subspace indexing, providing soft and hard masking
strategies with online adaptation capabilities. We demonstrate
that goal-specific masks can achieve 25-50% sparsity while
maintaining or improving retrieval accuracy across RF signal
processing and speech recognition tasks. The framework supports
standardized JSON outputs, mask diagnostics, and provides
interpretable explanations of feature importance. Experiments
show that our method outperforms PCA-based dimensionality
reduction by 15-30% in retrieval accuracy while providing 2-4×
speedup in query processing.