Conversational Shopping Agent

Conversational Shopping Agent is a retrieval-augmented conversational agent built for the TikTok TechJam 2026 hackathon. A simulated shopper describes one specific product from a frozen 50,000-item Amazon clothing catalog a few facts at a time; over at most 10 turns, the agent has to return ten ranked product IDs and may ask one clarifying question, and is scored on a weighted blend of Hit@10, Mean Reciprocal Rank, and how early the correct product was first surfaced. Each turn runs a three-stage pipeline. Dialogue accumulates every disclosed fact into a single running query and detects when the shopper changes their mind mid-conversation. Retrieval fuses BM25 sparse search over the catalog with dense embedding similarity from a bge-small-en-v1.5 bi-encoder using reciprocal rank fusion, then reorders the top candidates with a listwise LLM reranker built on RankGPT's permutation-generation technique. Control validates the ranked IDs against the catalog, pages through unseen items first, and always asks a broad clarifying question to draw out as much information as possible each turn. This hybrid retrieval and reranking pipeline lifted the score from a 0.10671 BM25-only baseline to 0.828623, with the listwise reranking step responsible for most of that gain. Working in a two-person team, I owned the Retrieval stage end-to-end: the BM25 index, dense retrieval, reciprocal rank fusion, the offline embedding index build, and the LLM reranking integration.