This paper proposes a novel retrieval framework, SPARC (Soft Probabilistic Adaptive Retrieval Model via Codebooks), to address three key challenges of multi-interest modeling in practical recommender systems (RS): 1. invariant interests extracted from predefined external knowledge, 2. over-exploitation strategies focused on matching existing interests, and 3. lack of novel interest discovery. SPARC utilizes a Residual Quantized Variational Autoencoder (RQ-VAE) to construct a discrete interest space, which is then trained alongside a large-scale recommender model to mine behavior-based interests that dynamically evolve and reflect user feedback. Furthermore, a probabilistic interest module that predicts the probability distribution over the entire dynamic discrete interest space enables an efficient "soft search" strategy during online inference, shifting the paradigm from passive matching to active exploration and effectively facilitating interest discovery. A/B testing on an industry platform with tens of millions of daily active users yielded significant results, including a +0.9% increase in user watch time, a +0.4% increase in page views (PV), and a +22.7% increase in PV500 (new content reaching 500 PVs within 24 hours). Offline evaluations using the Amazon Product dataset also consistently showed improvements in metrics such as Recall@K and NDCG@K.
