Cloudflare announced AI Search general availability on October 1, adding native image embeddings, optical character recognition for scanned PDFs and larger supported files. For people assessing a search product, the consequential change is how the system represents visual material before retrieving it. Cloudflare’s announcement
The service can now use Qwen3-VL-Embedding to represent images directly. Cloudflare also describes a caption-based path for text-only embedding choices: an image is converted into a description, and that text becomes the searchable representation. These routes expose different information to retrieval. A caption can summarize an image, while a native image representation can retain visual relationships that a short description may omit.
That distinction is an editorial interpretation of the product design, not a measured claim that one route wins every search. The announcement does not supply an independent comparison for a reader’s document collection. Questions about the images, such as layout, depicted objects or written text, may place different demands on the representation. A product’s support for images alone does not resolve those differences.
Scanned-PDF support adds another layer. Cloudflare says optical character recognition is available to all accounts and billed as image processing. The text and PDF file limit rises from four to ten mebibytes. These changes broaden what can enter the collection, but file acceptance does not guarantee that every detail is extracted or that a generated answer is accurate.
Commercial timing is also explicit. Billing is scheduled to begin November 1, 2026. Published free allowances include five million ingestion tokens, ten gigabytes of storage, and separate allowances of 1,000 semantic queries and 1,000 full-text queries. Readers estimating usage should preserve the distinction between those query types and account for the representation they choose.
My inference is that general availability moves the decision from access to fit. The relevant evidence is whether important material survives ingestion and appears in useful search results for the intended collection. Native image retrieval and scanned-document support create more options now. Video and audio support remain future work, and should stay outside any claim about the current release.