RAG-Based Volleyball Tactical Review from Play-by-Play Data: Serve-Receive-Attack Chain Analysis

Abstract

Play-by-play volleyball logs preserve the order of contacts that box scores collapse, but automated tactical review depends on accurate rally reconstruction and dependable evidence retrieval. This study evaluated an end-to-end retrieval-augmented generation pipeline on the 2025 NCAA Division I women's volleyball season. Canonicalization reduced 7,505,822 raw events to 5,654,956 by removing 1,684 verified duplicate full-match logs. A serve-anchored parser reconstructed 823,508 rallies, of which 821,923 (99.81%) had a valid sequence and point winner. Across the league, side-out was 57.13%, organized receive-set-attack occurred on 76.51% of receive opportunities, and the source-labeled first-ball kill rate was 29.26% (28.88% under the strict complete-chain funnel). Match winners exceeded losers by 12.02 percentage points in side-out and 7.56 points in first-ball kills. Pressure rallies showed a 1.17-point unadjusted side-out increase, but the team-, rotation-, and score-stratified odds ratio was 1.012 (p = 0.198). The pipeline created 60,852 match-team-rotation evidence cards and evaluated 270 retrieval queries. Structured hybrid retrieval achieved 99.63% Recall@1 and 100% Recall@5, compared with 30.37% and 61.48% for TF-IDF. Across 36 tactical reviews, chain-and-rotation retrieval produced 53.33% evidence coverage, 88.33% tactical specificity, a 68.10 composite quality score, and 100% numeric precision. In the championship final, Texas A&M's 71.67% side-out rate was driven by free points and extended-rally conversion rather than superior first-ball finishing. The workflow converts volleyball event streams into traceable, rotation-specific tactical narratives.