Alibaba Tongyi's open-source full-modal knowledge base RAG framework, VimRAG.

Alibaba's Tongyi Lab has launched the VimRAG framework, which replaces linear context stitching with dynamic directed acyclic graphs (DAGs). Through a branching trial-and-error mechanism, visual energy allocation strategy, and graph-guided optimization (GGPO), it achieves accurate retrieval and association reasoning across text, images, and videos. Experiments show that the model achieves an accuracy of 50.1% on the Qwen3-VL-8B model, significantly outperforming traditional solutions.