Interactive video event retrieval must balance retrieval quality, latency, and memory use. We present a fully local online retrieval pipeline using Qwen3-VL-Embedding-2B as a shared backbone for keyframes, video shots, speech transcripts, and captions. Evidence is indexed offline, while query encoding, retrieval, rank fusion, and duplicate suppression execute locally. We compare eleven configurations on 76 organizer-provided test cases with team-verified ground truth. Retrieval quality is evaluated retrospectively, using ground truth to select the best-performing query for each case and configuration from a shared bank of 174 single-event formulations derived from the task descriptions. Under this protocol, direct shot retrieval achieves the highest observed aggregate scores, including 78.95% Recall@20. In a separate fixed-query benchmark on an NVIDIA GeForce RTX 3070 Laptop GPU, this configuration achieves a median warm-query latency of 1.91 seconds and peak process resident memory of 4.22 GiB. Configurations incorporating speech transcripts or captions recover some cases but lose others, without consistently improving aggregate retrieval quality. These results characterize the evaluated configurations and highlight trade-offs between evidence selection, retrieval effectiveness, and computational cost.