deepvideodiscovery

Deep Video Discovery (DVD) is the official code release for a research paper on agentic search and tool use for long-form video understanding. The system is designed to answer complex questions about very long videos by treating segmented clips as an exploration environment. An LLM-driven agent plans its approach, selects multi-granularity tools to inspect relevant portions of the video, summarizes observations, reflects on them, and builds a final answer. The repository provides a Python implementation with dependency installation through pip and configuration in config.py. The included local_run.py example accepts a YouTube URL and a natural-language question. It also supports OpenAI and Azure OpenAI APIs, a Gradio demo, automatic subtitles, and a lite_mode that uses only subtitles, which the project describes as suitable for YouTube podcast analysis. For reproduction and evaluation, the project includes REPRODUCE.md for LVBench and links to caption data for LVBench, Video-MME, LongVideoBench, and EgoSchema, including an LVBench variant with transcripts. The README reports state-of-the-art results on several long-video benchmarks using OpenAI o3. Planned work listed in the repository includes an MCP server and release of evaluation trajectory data. The code is released under the MIT license.