Search mixed PDFs and images at once with Gemini

Extract data from complex documents containing both text and charts without OCR complexity. Gemini's multimodal file search now indexes PDFs and images in a single vector database, letting you query financial reports alongside product photos or legal docs with diagrams-answering questions like "Which charts show Q3 revenue growth?" with full citations.

Enterprise data is messy: financial reports contain embedded charts, contracts have handwritten annotations, technical specs mix diagrams with prose. Until recently, retrieving information from these mixed-media documents required converting images to text through OCR-a slow, error-prone bottleneck. Gemini's File Search tool changed this in May 2026 by adding native multimodal support powered by the Gemini Embedding 2 model.

The approach is simpler than traditional RAG pipelines. You create a File Search store specifying the embedding model (for multimodal, use gemini-embedding-2), upload your files (PDFs, PNGs, JPEGs, WebP), and the system automatically indexes both text and visual content into a single shared vector space. A professional searching their earnings-release library no longer needs separate text and image databases-one query like "Show me charts comparing year-over-year revenue growth" retrieves relevant visualizations from PDFs without pre-processing.

For practical work: lawyers reviewing contract amendments alongside regulation diagrams, analysts extracting patterns from quarterly reports with embedded charts, or product teams analyzing competitor product photos mixed with spec sheets all benefit from this unified search. Metadata filtering lets you narrow results (e.g., only finance department documents from 2025), and page citations show exactly where each answer came from.

The cost structure favors developers: storage and query-time embeddings are free; you pay only for indexing and tokens. Setup requires the Gemini API (Python, JavaScript, or Java SDKs supported), but the learning curve is gentle-the File Search documentation includes working code examples for immediate experimentation.