Cloudflare made AI Search generally available with image retrieval and support for scanned PDFs. The new pipeline embeds an image’s visual content alongside its caption, allowing searches to use details the caption leaves out. It also extracts text from scanned documents before indexing. Developers can retrieve relevant document chunks directly or pass them to a chosen chat model to generate an answer.
Cloudflare made AI Search generally available on 1 October, adding visual retrieval and support for scanned PDFs to the managed search service. Developers can use it to search their own documents and websites. The indexing pipeline brings together Workers AI, Vectorize, R2 and Browser Run, Cloudflare services that prepare and retrieve the customer’s material.[1], [2]
Previously, image search relied on matching the caption generated from a picture. The new version also sends the image’s pixels to an embedding model. Qwen3-VL-Embedding maps text and pictures into a shared representation space, allowing a text query to be compared with visual material. When developers select a text-only model, ToMarkdown still creates the captions used to represent images.[1]
Optical character recognition extracts text from scanned PDF pages before the document is split into chunks and indexed. Optional query rewriting can precede retrieval. During retrieval, vector and keyword searches run in parallel; their results are combined and reranked. Applications can obtain relevant chunks directly or send them to a chosen chat model for answer generation. Cloudflare announced billing from 1 November, while identifying audio and video support as subsequent work.[1]
