Rose Khairoalsendi

تخصص الذكاء الاصطناعي · دفعة 2026

Generative Late-Interaction Embeddings For Visual Document Retrieval

Late-interaction models such as ColPali achieve strong visual document retrieval performance but require storing many token embeddings per page. We propose compact generative codes that capture each page’s low-dimensional token manifold. The project has two central components: efficient retrieval and on-demand vector generation. The codes can be scored directly using standard MaxSim for first-stage retrieval, then used to generate richer token representations on demand for more accurate reranking—without permanently storing those vectors. Anchored by cluster centroids and refined using the complete token set, the codes are trained through knowledge distillation from a frozen retrieval encoder.

اقرأ التقرير النهائي (PDF)