Project description Shopping Copilot is a conversational product-search agent for the Amazon Clothing, Shoes & Jewelry catalog. It helps shoppers move from vague requests to relevant product recommendations through multi-turn conversation. The agent uses two interaction routes:
- Browsing route: for exploratory requests such as “I’m looking for something for a wedding.”
- Buying route: for specific requests containing details such as colour, budget, size, material, or product type. It tracks session memory across messages, extracting preferences including category, colour, material, size, budget, and use case. When a shopper changes their mind—for example, “Actually, not boots; I need a handbag”—the agent updates its stored state and replaces conflicting preferences. Products are searched locally from the provided catalog using an in-memory SQLite FTS5/BM25 index. The agent asks early clarification questions to gather missing details, then returns up to 10 matching product IDs for evaluation. On the public development set, the current version achieves: Hit Rate@10: 0.855 MRR: 0.532 MTTC: 3.65 Development tools
- VS Code
- Python 3
- Git and GitHub
- TechJam local evaluator APIs used No external APIs or paid LLM APIs were used. The project runs locally and in memory. Libraries and frameworks
- Python standard library
- sqlite3 with FTS5 for BM25 full-text search
- json for catalog/session data loading
- re for rule-based preference extraction
- pathlib for file handling Dataset and assets
- TechJam’s frozen 50,000-product Amazon Reviews 2023 Clothing, Shoes & Jewelry catalog
- TechJam’s 200 public labelled development sessions
- The official TechJam starter agent, API contract, and local evaluator
Log in or sign up for Devpost to join the conversation.