Cached Descriptive Asset Files (CDAF): A Sidecar Format for Token-Efficient Video Understanding in Agentic Pipelines

Multimodal language models can answer questions about video, but agentic workflows often pay the full video-processing cost every time the same asset is reused. We present Cached Descriptive Asset Files (CDAF), an open, model-agnostic sidecar convention that persists a timestamped description alongside a video and binds that description to the exact video bytes using SHA-256. An agent verifies freshness and reads a few hundred text tokens rather than repeatedly exposing the model to the video.

In a reproducible synthetic benchmark containing four 10–12 second videos and 20 objective questions per condition, CDAF-mediated question answering achieved 20/20 accuracy, compared with 19/20 for direct video analysis, while reducing mean prompt tokens from 3,066 to 303 (10.1×) and mean latency from 3.46 to 2.24 seconds (35.2%). Conservative accounting places the one-time generation break-even at approximately 1.3 downstream questions per video.

We specify the file format, freshness semantics, agent decision policy, and evaluation harness; analyze how CDAF converts footage selection and cut-list generation into text retrieval; and identify failure modes involving omitted detail, malicious descriptions, and synthetic-benchmark realism. The result is a deliberately simple systems primitive: video understanding becomes a reusable, verifiable text asset rather than a repeated multimodal inference cost.