Music search at the scale of Amazon Music presents a unique challenge: queries frequently deviate from indexed metadata due to misspellings, transpositions, and phonetic variations, yet the retrieval system must operate under strict millisecond-level latency constraints. Our existing learning-to-retrieve system, the High Confidence Index (HCI), learns query-entity associations from customer behavior, relying on continual "exploration" to choose candidates. Traditional n-gram matching enables this exploration but suffers from poor semantic robustness and high noise, limiting the system's ability to learn from long-tail queries. In this work, we present a robust neural sparse retrieval system designed to maximize exploration efficiency. We adapt a state-of-the-art inference-free sparse retrieval architecture to the music domain, combining it with an effective domain-specific granular subword tokenization strategy. Our approach utilizes short-length token constraints (max 3 chars) to enforce the learning of surface-form robustness over lexical memorization. By pre-computing the neural embeddings and term expansions during the offline indexing phase, online processing is reduced to minimal tokenization and IDF weighting, achieving effectively zero latency overhead for query encoding. Evaluations on a 6M-document production corpus show an aggregate 91.4% recall@10 (vs. 57.7% for trigrams) at comparable throughput. Simulation of the HCI feedback loop demonstrates improved exploration efficiency, with +0.8% higher stabilized recall than production trigrams. Ablation studies indicate that our sparse training methodology drives the performance gains, while domain-specific pretraining provides a cost-effective alternative to large-scale general-purpose pretraining.

This work presents the search and discovery architecture of Amazon Music, a highly efficient system designed to retrieve relevant music content for users. The architecture consists of three key stages: indexing, retrieval, and ranking. During the indexing stage, data is meticulously parsed and processed to create a comprehensive index that contains dense representations and essential information about each document (such as a music or podcast entity) in the collection, including its title, metadata, and relevant attributes. This indexing process enables fast and efficient data access during retrieval. The retrieval stage utilizes multi-faceted retrieval strategies, resulting in improved identification of candidate matches compared to traditional structured search methods. Subsequently, candidates are ranked based on their relevance to the customer’s query, taking into account document features and personalized factors. With a specific focus on the podcast use case, this paper highlights the deployment of the architecture and demonstrates its effectiveness in enhancing podcast search capabilities, providing tailored and engaging content experiences.