← Back to feed

Google has updated the Gemini API File Search tool with three capabilities that expand what developers can build on top of it: multimodal retrieval, custom metadata, and page-level citations. The announcement landed on May 5 and surfaced on Hacker News front page this weekend.

What Changed

Multimodal support. File Search can now retrieve across images and documents in the same corpus, not just text files. A query that might retrieve a paragraph from a PDF can now also surface a relevant chart, diagram, or screenshot from the same document or a different file in the collection. For mixed-media knowledge bases — technical manuals, research reports, product catalogs — this removes the requirement to build separate retrieval pipelines per content type.

Custom metadata. Developers can now attach metadata fields to files and use them as filters during retrieval. A legal document store can be filtered by jurisdiction, date range, or document type before the model generates a response. A customer support corpus can be scoped by product line or region. Without this, RAG pipelines had to do post-hoc filtering or run separate indexed views. The in-API filter collapses that into a single call.

Page-level citations. Responses can now cite specific pages within a retrieved document rather than the document as a whole. For enterprise deployments where provenance determines whether a legal, compliance, or finance team will trust the output, document-level attribution is not sufficient. Page citations are the delta between a demo and a production deployment.

The Competitive Position

OpenAI has offered file search and retrieval through the Assistants API since late 2023, with periodic improvements to citation quality. Anthropic handles RAG largely through the model’s extended context window rather than a managed retrieval layer. Google is positioning the Gemini API as the developer-first RAG infrastructure with managed retrieval, multimodal indexing, and citation infrastructure included.

The three capabilities together address the most common failure modes in enterprise RAG adoption: mixed-media corpora that resist text-only pipelines, retrievals that return irrelevant files from adjacent categories, and answers that can’t be traced back to a page. Whether the implementation quality matches the spec is a developer question. The API is available now.

Numbers

• Multimodal support: Images + documents in the same retrieval corpus
• Custom metadata: Per-file tags, filterable at query time
• Page citations: Response attribution to specific pages, not just document-level
• Availability: Gemini API, generally available
• Published: May 5, 2026 (Google DeepMind)