Text-to-audio retrieval is a fundamental task in acoustic signal processing. Currently, mainstream approaches primarily employ a dual-tower architecture, independently encoding text and audio while performing similarity score matching. However, these methods struggle to maintain a uniform embedding space and the latency of score matching increases with corpus size. In this light, we propose GTA to move towards a generative text-to-audio retrieval paradigm. Specifically, we utilize a multi-scale audio tokenizer to embed audio semantics into identifiers, while incorporating a dual-alignment strategy to ensure consistency between audio and text semantics. Furthermore, we implement curriculum learning to bridge the gap between training and inference, guiding the model toward precise token-level generation. Extensive experiments demonstrate the effectiveness and robustness of GTA, further validating the feasibility of the generative text-to-audio retrieval paradigm.