Word spotting enables archaeologists, historians, and content moderators to retrieve regions of interest from document images based on specific queries. While mature methods exist for handwritten documents in languages like English, Chinese ancient documents pose greater challenges due to issues such as paper degradation, uneven coloring, a large character set, and the presence of rare or unseen characters. To address these challenges, we propose an end-to-end segmentation-free word spotting method based on Hierarchical Decomposition Embedding (HDE), which simultaneously predicts word bounding boxes and embeddings in a single stage. By learning mappings between character regions and hierarchical embedding spaces—instead of relying on closed-set classification—the model effectively handles unseen character queries. Experiments on the MTHv2 dataset demonstrate a state-of-the-art MAP of 80.84%. This work offers a novel approach for the preservation and retrieval of Chinese ancient documents.