Kurly Shares Case of Automating Review Analysis by Calling Gemini Pro from BigQuery

Kurly's Data Service Development Team unveiled a method of automating review analysis work by directly calling Gemini Pro within the BigQuery environment. Previously, building a separate deep learning model was required, but after integrating Vertex AI and utilizing SQL's ML.GENERATE_TEXT function, implementation time and cost were greatly reduced.

Key Use Cases

• Review Summarization and Promotional Copy Generation: Combines reviews per product to instantly generate a one-word summary or marketing copy.
• Keyword Extraction and Category Classification: Extracts key keywords from review text or classifies them into predefined categories such as delivery and quality.
• Sentiment Analysis: Quantifies the sentiment of reviews, such as by scoring them on a numeric scale of 0 to 10.

Reliability Assurance and Large-Scale Processing Strategy

To address the hallucination issue characteristic of LLMs and inconsistencies in output format, responses were forced into JSON format, with format instructions clearly included in the prompt. Additionally, when processing large volumes of data, a batch technique grouping reviews in units of 100 was applied to improve efficiency. Through the RAI (Responsible AI) feature, violations of safety attributes can be checked, and by checking the finish_status value, the reason for execution termination—such as generation completion or token overflow—can be tracked.