Understanding new Gemini model inference, tuning, grounding and safety settings in BigQuery

The exponential growth in unstructured data from digital devices and platforms necessitates advanced tools for analysis. BigQuery, Google’s AI-ready cloud data platform, integrates with Vertex AI to leverage generative AI models for unstructured data processing. This integration allows the use of models like Gemini for tasks including text summarization and sentiment analysis. BigQuery also supports fine-tuning models using LoRA techniques, which is useful when prompt engineering is insufficient. Recent updates include the addition of Gemini 1.5 models, which enhance NLP, vision tasks, and new capabilities like audio transcription and PDF summarization. The ML.GENERATE_TEXT SQL function now supports grounding with Google search and customizable safety settings to ensure responsible AI outputs. Additionally, BigQuery extends support for Gemini 1.0 model tuning and evaluation, enabling tailored AI capabilities. Users can create remote models representing Vertex AI Gemini endpoints and process unstructured data with object tables in BigQuery. Grounding and safety settings offer detailed control over the AI responses, ensuring accuracy and adherence to defined safety thresholds. Fine-tuning with LoRA for Gemini models allows precise model behavior customization for specific applications.