Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Thomas Austin

Second Advisor

Katerina Potika

Third Advisor

Amith Kamath Belman

Keywords

dish-to-grocery mapping, structured prediction, large language mod- els, schema-guided generation, index-constrained product selection, entity matching, retailer catalog retrieval, agentic orchestration

Abstract

Consumers define what they want to prepare as dishes, while retailer systems organize groceries through product-specific store identifiers. This mismatch makes it difficult to map consumer-defined dishes to grocery products using keyword search or arbitrary language generation. I proposed SmartEcom, a prototype system that accepts dish queries in the user’s native language, converts them into standardized ingredient lists, and maps those ingredients to actual product offerings from Kroger and Walmart. SmartEcom’s major contribution is an index-constrained matching stage. After candidate products are identified by a retailer API, a large language model selects among them by choosing their indices instead of generating product identifiers. This makes the selected products deterministic, traceable/auditable, and resistant to hallucinated UPCs (Universal Product Codes). This report also evaluates how OpenAI generative models produce ingredient lists during extraction. Both GPT-5 and GPT-4o generate relevant ingredient lists, but GPT-5 tends to produce longer lists and is significantly better with regional naming conventions and lexical variations. Although overall pass rates are similar, GPT-5 provides improved recipe comprehension and ingredient completeness.

Download

Available for download on Saturday, May 22, 2027

Share

COinS