Caption Pipe

This project is a Dockerized video understanding and caption generation pipeline built for short-form clips. Given a task list of video URLs, it downloads each video, normalizes the media into a stable analysis format, extracts representative frames across the video, and divides the clip into temporal segments for structured reasoning. Each segment is analyzed with a vision-language model using the selected frames, optional transcript context, and accumulated continuity memory from earlier segments. Those structured segment outputs are then merged into a deterministic per-segment ground-truth layer and carried into a final global factual summary step, which remains image-backed to preserve direct visual grounding. The factual summary serves as the source of truth for final caption generation, allowing the system to produce multiple caption variants such as formal, sarcastic, and humorous styles without inventing unsupported details. The pipeline is designed to be debuggable and operationally practical: it preserves intermediate artifacts, logs each stage, tolerates partial failures, and uses bounded concurrency for subprocesses and model calls. The result is a robust, evidence-first workflow for turning raw video clips into high- quality, style-controlled captions.