Imagine effortlessly searching through countless hours of video footage to find precisely what you need. A new research project, VOCAL-UDF, is tackling this challenge by building a "self-enhancing" video data management system powered by large language models (LLMs). Traditional video search tools struggle with complex queries like "a motorcycle swerving near a silver Subaru and then colliding with it." These queries involve multiple steps, object recognition, relationship understanding ("near," "colliding"), and attribute identification ("silver"). LLMs, while powerful, aren’t designed to handle these multi-faceted video searches efficiently. VOCAL-UDF introduces a novel approach: it automatically generates the missing "building blocks" needed to understand these queries. It breaks down complex requests into smaller, manageable sub-tasks. If a necessary component is missing – like a module to identify "silver" cars – VOCAL-UDF automatically builds it using LLMs, learning and expanding its abilities over time. This system combines the power of LLMs with two types of user-defined functions (UDFs): program-based UDFs (Python code) and distilled-model UDFs (compact AI models). Program-based UDFs handle tasks like calculating distances between objects, while distilled-model UDFs are trained on the fly to recognize more nuanced visual concepts, like a "swerve." To further enhance accuracy, VOCAL-UDF generates multiple possible interpretations of a concept (e.g., different definitions of "near") and uses active learning to efficiently identify the most accurate one with minimal user feedback. Tested across diverse video datasets (traffic surveillance, daily activities, synthetic animations), VOCAL-UDF significantly boosts search accuracy. This research opens exciting possibilities for a future where querying videos is as easy as asking a question, paving the way for powerful video analytics tools in numerous applications. While still in early stages, tackling the challenge of building compact, efficient models for visual reasoning remains a key next step for realizing the full potential of this technology.