Cloudflare Launches General Availability for AI Search with Enhanced Multimodal Capabilities
Cloudflare's AI Search is now generally available, introducing native image embedding, OCR for PDFs, and support for larger files to enable richer multimodal search experiences.
Cloudflare has announced the general availability of its AI Search service, a comprehensive managed pipeline that integrates Workers AI, Vectorize, R2, and Browser Run. This move signifies a maturation of the product, which has been in use for over a year, powering various search applications from internal documentation to public-facing websites, including Cloudflare's own blog and developer documentation.
The most significant enhancement accompanying the general availability is the expansion of multimodal search capabilities. AI Search now supports native image embeddings, allowing users to search for content based on visual similarity rather than solely relying on text captions. This feature preserves both the visual signals within an image and its associated text, leveraging Matryoshka Representation Learning (MRL) to maintain efficient embeddings for fast and manageable storage and retrieval.
Previously, image search was a more rudimentary process involving object detection and text captioning before embedding. The new implementation allows for direct image embedding by multimodal models like Qwen3-VL-Embedding, enabling searches that capture fine visual details such as texture, composition, and color. Even for text-only embedding models, AI Search can convert query images to text using ToMarkdown, providing a baseline multimodal search functionality.
Beyond images, AI Search now offers enhanced support for PDFs and larger files. The service can now ingest text files and PDFs up to 10 MiB, a substantial increase from the previous 4 MiB limit. Crucially, it introduces Optical Character Recognition (OCR) for scanned PDFs, enabling the extraction of text from image-based documents before chunking and embedding, thereby making previously unsearchable documents accessible.
With the transition to general availability, Cloudflare is also introducing billing for AI Search, set to commence on November 1, 2026. The pricing model is designed for predictability, based on content ingestion, data storage, and query volume. The core processing tasks, including parsing, chunking, embedding, and reranking, are included in the service, with no additional costs for instance hours or capacity units. A generous free tier remains available across all Workers plans.
The pricing structure includes a base ingestion rate per token, storage costs based on index size, and query fees differentiated by semantic or full-text searches. Cloudflare has also slightly adjusted the free monthly allotment, providing 1,000 semantic and 1,000 full-text queries separately, rather than a combined pool.
This general availability marks a significant step for Cloudflare in making advanced AI-powered search accessible and practical for developers. The enhanced multimodal capabilities, including native image retrieval and OCR for scanned documents, broaden the scope of data that can be effectively searched and analyzed, positioning AI Search as a versatile tool for a wide array of applications.
The service's ability to handle richer data formats and provide more nuanced search results addresses a growing need for sophisticated information retrieval in an increasingly visual and complex digital landscape. The predictable pricing model further encourages adoption by businesses and developers looking to integrate advanced AI search functionalities into their platforms.