Google Makes Gemini API File Search Multimodal With Page-Level Citations
Google's Gemini API File Search now indexes images alongside text, ties responses to original page numbers, and ships free storage and query embeddings — turning verifiable RAG into a one-call developer primitive.
Gemini API File Search Now Sees What Your Documents Show
Google announced on May 5, 2026 that the Gemini API File Search tool now supports multimodal retrieval-augmented generation, custom metadata filters, and page-level citations. For developers building production RAG pipelines on top of the Gemini API, this is the kind of platform update that quietly removes a category of work — the bespoke pipeline of OCR, vector store provisioning, and citation bookkeeping that used to sit underneath every reasonable RAG application.
The feature most worth highlighting is multimodal indexing. The Gemini API File Search tool now processes images and text together, powered by the Gemini Embedding 2 model, which natively understands image data. That means a single search store can hold PDFs, DOCX files, slide decks, JSON, Jupyter notebooks, Markdown — and PNG and JPEG images up to 4K resolution — in one place, and a query can match against the visual content of charts, diagrams, and product photos as readily as it matches against the surrounding prose.
Page-Level Citations Make RAG Verifiable
The second meaningful change is that File Search now captures the page number for every piece of indexed information and ties model responses directly back to the original source. Page-level citations turn a generated answer into something the calling application can audit — a developer can render a Gemini API answer alongside a link to the exact page of the underlying PDF, which is the citation pattern that enterprise customers have been waiting for to feel comfortable shipping RAG into client-facing surfaces.
Custom Metadata Filters Let Agents Reason About Scope
The third update is custom metadata filtering. Indexed files can now carry developer-defined tags, and queries can scope retrieval to subsets of a corpus — by document type, customer, region, version, or any other axis a developer cares to define. That is the building block agentic systems need to reason about the scope of their answers without the agent having to do filtering work in-context.
A Pricing Model Designed for Production
The economic structure of the new Gemini API File Search release is unusually friendly to production deployments. Storage is free. Query-time embeddings are free. Customers pay only for the initial indexing embeddings and standard Gemini input and output tokens. That cost shape rewards long-lived knowledge bases — the kind of corpus where you index once and then query many times — and it removes the cost-per-query anxiety that has historically constrained RAG architectures from running at scale.
What This Means for Multimodal RAG Architectures
The combination of native multimodal embeddings, page-level citations, and metadata filtering means the Gemini API File Search tool is no longer a primitive that developers wrap with their own RAG pipeline — it is the RAG pipeline. For teams that have been maintaining custom OCR, custom chunking, custom citation layers, and custom vector stores around a Gemini call, the May 5 release is an invitation to delete a meaningful amount of code.
The supported file types list is broad enough to cover most real-world enterprise corpora: PDF, DOCX, TXT, Excel, CSV, JSON, SQL, Jupyter notebooks, HTML, Markdown, and the new addition of PNG and JPEG up to 4K. Combined with the multimodal embedding model, that lets a single index hold the source documents, the slide decks, and the chart images that together describe a domain — and lets an agent reason across all three modalities in a single retrieval call.
The Bigger Multimodal RAG Picture
Multimodal RAG has been the hard problem in agentic AI for the past year — most enterprise documents are not pure text, and the cost of converting charts and diagrams into prose for indexing has been a structural ceiling on how useful RAG could be. Google's May 5 release tackles that ceiling directly. It is also a strong signal about where Google sees the agentic developer surface heading: toward primitives that turn multimodal verifiable retrieval into a single API call, rather than a bespoke architecture every team has to assemble themselves.
Sources: Google blog, May 5, 2026; Google Developers Blog, May 5, 2026; Google AI for Developers documentation, May 2026.