High-fidelity reasoning and question answering in long-form video require compact, multimodal representations. Existing methods compress video segments into deterministic text captions, an approach that is inherently lossy and discards the natural semantic ambiguity of visual data. To resolve this, we propose a probabilistic tree of embeddings framework that models video segments as multivariate distributions directly in the visual embedding space and organizes them in a tree structure, allowing efficient segment retrieval while preserving vital visual uncertainty for downstream reasoning tasks. To realize this framework at scale, we must train advanced generative models to efficiently map raw video chunks into these complex probabilistic distributions. Therefore, we are requesting HPC resources to support this computationally intensive training phase, which requires processing millions of high-dimensional frames across massive, unlabelled video datasets. Once trained, this framework will eliminate traditional inference bottlenecks, providing a highly scalable foundation for complex multimodal reasoning.