With this update, developers can now integrally process image data in addition to text. By leveraging the Gemini Embedding 2 model, it is now possible to search by specifying visual atmosphere or style using natural language.
Product LaunchesGoogle DeepMindGeminiGemini API File Search
Gemini API File Search Now Supports Multimodality, Adding Metadata and Page Citation Features
This article is a translation. Read the Japanese original
Additionally, a feature to attach custom metadata to unstructured data has been added. By assigning key-value labels such as department or status and filtering during queries, noise from unrelated documents can be reduced.
Furthermore, a function has been introduced that allows the source of an answer to be traced back to a specific page number. This enables the explicit identification of where an answer derived from a large volume of PDFs originated, contributing to improved fact-checking and reliability.
Source: Gemini API File Search is now multimodal (HN 156pt, 46 comments) (HN Search (backfill), 2026-05-10)