Dialogue-Aware Conversational Music Recommendation via Multi-Modal Retrieval and Weighted RRF

The ACM-RecSys 2026 Music Conversational Recommendation Challenge (Music-CRS) bridges natural language processing and recommender systems, requiring systems to engage in multi-turn dialogues and recommend relevant tracks from a large catalog. In this paper, we present Team Sargam’s solution, a robust hybrid retrieval and reranking pipeline. Our approach combines deterministic artist and album isolation heuristics with multi-seed multi-modal FAISS similarity search (spanning collaborative filtering, audio, and visual embeddings), and sparse BM25 lexical retrieval. We aggregate these disparate streams using a tuned Weighted Reciprocal Rank Fusion (RRF) mechanism. For high-specificity dialogue sessions, we introduce an exact-target extraction step using Large Language Models (LLMs) to isolate the exact track requested. Finally, we employ a constrained LLM to generate diverse, context-aware responses. Our lightweight, training-free pipeline achieved a composite score of 0.4482, securing 9th place in the Academic Track and 25th place overall. Our code and recovered test datasets are publicly available at https://github.com/this-AR/sargam-music-crs.