In recent years, increasingly diverse customer values and lifestyles in housing selection have underscored the need for flexible floor plan retrieval systems that reflect subjective perspectives such as storage sufficiency, circulation efficiency, and spatial connectivity. However, conventional retrieval approaches based on labeled graphs struggle to represent fine-grained spatial structures and room attributes, leading to limited accuracy and flexibility. To address these limitations, we model each floor plan as an attributed graph that encodes both inter-room connections and room-specific attributes, and we define twelve customer-oriented conditions as the presence or absence of characteristic subgraphs. We then propose a vectorization framework that combines ResNet-18 image features with these condition vectors in a contrastive-learning paradigm. The resulting embedding space places floor plans with similar conditions in close proximity. Experiments show superior classification performance on several conditions, confirming the effectiveness of the proposed approach.