Google is rolling out three updates to the File Search tool in its Gemini API. First, multimodal retrieval: powered by the Gemini Embedding 2 model, developers can upload images and text into a single knowledge base and search everything using natural language. You could, for example, find images that match a specific visual style or emotional tone. Second, custom metadata filtering: when uploading files, you can attach key-value tags like department: Legal, and then filter queries by those tags to narrow results. Third, page-level citations: the model will tell you exactly which file and page number a piece of information comes from, making it easy to jump in and verify.
File Search is Google's fully managed RAG (retrieval-augmented generation) system built into the Gemini API, handling file storage, chunking, vectorization, and context injection automatically. The embedding generation during storage and query is free; you only pay during initial indexing, at $0.15 per million tokens.