phi-subgroup — Video Search Task

Is this run manual or automatic?: automatic
Describe the retrieval model used.: We extract SigLIP2-base-patch16-naflex embeddings at 1 keyframe per second. Each user query is expanded to 100 variants using GPT-4.1-mini, and their text embeddings are averaged into a single query vector. Initial retrieval is done directly using SigLIP similarity, returning the top 2,500 candidates. Each candidate shot is then evaluated 10 times using Phi-3.5-Vision, and the scores are averaged. An overlapping subgroup sort is then applied during re-ranking to limit how far each result can move from its initial rank, and the top 1,000 results are submitted.
Describe any external resources used.: We use embeddings from SigLIP2-base-patch16-naflex, GPT-4.1-mini for query expansion, and Phi-3.5-Vision for LLM judgement.
Training type:: D