On May 5, 2026, Google announced a major upgrade to the Gemini API's File Search tool, officially supporting multimodal retrieval. The update is based on the Gemini Embedding 2 model, letting developers simultaneously understand and retrieve image and text content.
Previously, File Search only supported pure-text retrieval, requiring developers to use external tools to convert images to text before processing. Now, Gemini can directly read original images, finding visual assets via natural-language descriptions. For example, inputting a warm-toned advertising poster, the system can find matching items from the image library, no longer relying on file names or alt text. For enterprises needing to manage large amounts of visual material, this means a paradigm shift in retrieval logic.
The upgraded File Search also brings two practical new features: custom metadata filtering lets you attach key-value tags to files, and queries can directly limit scope, significantly reducing noise; page-level citations let every piece of information in the AI's answer be traced back to a specific page of the original PDF, improving transparency and making fact-checking easier.
Multimodal retrieval capability changes the design thinking of RAG systems — in the past, text and image indexes had to be handled separately, but now unified semantic retrieval is possible, simplifying architecture while improving recall quality. This is a pragmatic step by Google to land Gemini's multimodal capabilities into production tools, with File Search upgrading from finding documents to understanding content, and the way RAG works changing accordingly.