Dense document images often contain many fine-grained visual and textual entities whose relevance depends on a user query. Standard vision-language retrievers encode cropped regions with a single vector, which can mix distinct entity signals and obscure the evidence needed for fine-grained retrieval. We call this failure mode **Semantic Dilution** and quantitatively show that it degrades entity-level retrieval as a function of entity density. To mitigate it, we propose **SAGE**, a training-free framework that parses semantic entities from dense document images, represents them as hierarchical graph nodes with multi-vector embeddings, and retrieves query-relevant evidence through iterative entity-level subgraph matching. We also introduce **DEAR**, a dataset of 1,055 query–image pairs sourced from product detail pages, where each query requires retrieving and comparing multiple fine-grained entities from visually dense inputs across four question types of increasing complexity. Experiments show that SAGE substantially reduces semantic dilution and outperforms patch-level and OCR-based retrieval baselines on DEAR, achieving a Recall@3 of 0.849 and a generation score of 2.746 on multi-entity visual comparison queries. Our code is available at https://github.com/All4Nothing/SAGE.

Recently, many prior works have made their own agents generate more personalized and engaging responses using personachat. However, since this dataset is frozen in 2018, the dialogue agents trained on this dataset would not know how to interact with a human who loves “Wandavision.” One way to alleviate this problem is to create a large-scale dataset. In this work, we introduce the pipeline of creating personachatgen, which is comprised of three main components: Creating (1) profilegen, (2) Persona Set, and (3) personachatgen. To encourage GPT-3’s generation ability, we also defined a taxonomy of hierarchical persona category derived from social profiling taxonomy. To create the speaker consistent persona set, we propose a simple contradiction-based iterative sentence replacement algorithm, named CoNL. Moreover, to prevent GPT-3 generating harmful content, we presented two filtering pipelines, one each for profilegen and personachatgen. Through analyzing of personachatgen, we showed that GPT-3 can generate personalized dialogue containing diverse persona. Furthermore, we revealed a state-of-the-art Blender 90M trained on our dataset that leads to higher performance.