Gemini AI used for model training and multimodal retail recommendations

Developments in Google’s Gemini AI are being applied to both specialized data processing tasks and retail consumer experiences.

In a technical application of named-entity recognition (NER), Gemini 1.5 Pro was used to label thousands of Reddit comments regarding chef's knives to train a smaller, more cost-effective model. By using $9 worth of Gemini labels to fine-tune the open NER model GLiNER, researchers were able to achieve an 0.83 F1 score, significantly improving upon the zero-shot performance of the smaller model while reducing long-term operational costs.

In the retail sector, Gemini’s multimodal capabilities are being utilized to transform product discovery. Through Google Cloud applications, the AI can analyze visual scenes—such as a photograph of a living room—and combine that visual data with natural-language prompts to provide personalized recommendations for furniture, fashion, and home decor. This approach moves beyond traditional keyword searches by understanding colors, styles, and visual context.