Retrieving complex, tactical moments in specialized domains like badminton is a significant challenge that current Vision-Language Models (VLMs) are not equipped to handle. Trained on massive, generic datasets, VLMs learn pixel-to-text alignments but lack the domain-specific expertise to understand tactical intent or causality.
We propose a novel framework that bypasses this VLM alignment, instead leveraging the rich domain knowledge of Large Language Models (LLMs) to interpret structured event data. Our method first employs domain-specific computer vision tools to decompose videos into structured, textual game logs. We demonstrate that these game logs are a remarkably potent asset for retrieval, as semantic search on these logs alone dramatically outperforms state-of-the-art VLM-based systems.
To capture the causal reasoning that raw logs lack, we introduce a rigorous, multi-agent dialectic reasoning process where agents collaboratively debate the log, draft and revise a narrative, and verify its grounding to the original game log events. This ”Generate-then-Retrieve” framework provides a step-function improvement in retrieval accuracy. Our system is not only more accurate but also fully interpretable, providing human-readable, grounded explanations for every result.
