Cloudflare’s AI Search now supports native image embeddings and optical character recognition for PDFs, allowing users to search for visual details and scanned documents. The service uses Matryoshka Representation Learning to keep embeddings efficient and supports the Qwen3-VL-Embedding model for direct image queries. For text-only embedding models, AI Search converts query images to text using ToMarkdown, ensuring basic multimodal support across all models. This enables searches that are difficult to express precisely with words, such as finding a bird with similar markings or matching a screenshot by visual characteristics.

The service now accepts text files and PDFs up to 10 MiB, up from 4 MiB, and includes OCR for scanned documents. Cloudflare has also outlined its billing model, which will begin on November 1, 2026. Pricing is based on three components: content ingestion, data storage, and queries run. The company clarified that the free monthly allotment on Workers plans will include 1,000 semantic queries and 1,000 full-text queries, rather than a shared pool of 2,000 queries.

Cloudflare uses AI Search to power search on its own blog and developer docs. The service is designed to help developers estimate their bill before indexing content, with no instance hours, capacity units, or monthly minimums required upfront. The pricing structure aims to make it easy to size projects, with ingestion based on tokens and storage based on data size.