Search over the Visual World
Persistent Visual Memory, Layered Indexes, and Source-Grounded Evidence
Search over the visual world cannot be reduced to ranking video files. This report develops the infrastructure it does require — analyzer-defined scenes, persistent visual memory, capability-declared indexes, and evidence that stays playable at the source — and evaluates it against a commercial video-native retrieval engine over 9,834 natural-language queries drawn from four public datasets. The paper is published on arXiv (2608.08075).
The full report is a PDF. Open it to read all 33 pages, including the figures, tables, and the complete empirical study.
Open the paperTrouble viewing it here? Open the PDF in a new tab. Benchmark configurations and reproduction instructions are on GitHub, and the open-source Deep Search implementation of the stateful retrieval loop is public too.