BigQuery ObjectRef now GA: Build multimodal data platforms for conversational analytics entirely in SQL

For teams handling unstructured data like images and videos, BigQuery’s new ObjectRef feature eliminates the need for separate pipelines by allowing these files to be referenced as native columns within SQL tables. This integration reduces operational overhead by removing the post-hoc stitching of inference outputs back into the warehouse. Previously, structured and unstructured data required distinct workflows, but ObjectRef bridges this gap by treating unstructured files as first-class data within BigQuery’s analytical engine.

With ObjectRef, users can pass multimodal inputs directly to Gemini models via SQL or natural language queries, combining unstructured files with structured context such as metadata and timestamps. The feature supports fully serverless pipelines, enabling at-scale inference without separate compute clusters. Extracted features and annotations remain linked to the original unstructured objects, preserving nuanced context for conversational analytics and AI agents.

Customers have adopted ObjectRef to build production-ready pipelines for tasks like multimodal vector search, large-scale classification, and document feature extraction, all using standard SQL. The approach supports independent data lifecycles, allowing Cloud Storage objects and BigQuery tables to evolve separately without disrupting underlying data integrity.

Google Cloud introduced End User Credentials support for ObjectRefs in GA, enabling ad-hoc analysis using personal IAM identities, while also expanding metadata source compatibility and array processing capabilities. These updates streamline SQL-based multimodal pipelines, such as analyzing customer service calls by parsing audio files and extracting structured insights directly within BigQuery.