Cloudflare's AI Search reaches general availability

Cloudflare's AI Search is now generally available, more than a year after it launched. The product bundles Workers AI, Vectorize, R2 and Browser Run into a managed index and retrieval pipeline, so a team can put search over its own documents without assembling those pieces itself.

Images, PDFs and larger files

General availability arrives with wider input support. AI Search now embeds image pixels natively instead of relying on a generated caption, and adds optical character recognition for PDFs along with support for larger files. Cloudflare says it uses Matryoshka Representation Learning so that smaller embeddings keep useful detail while storage stays manageable.

One vector space for text and images

The new multimodal retrieval runs on Qwen3-VL-Embedding, which is available today. Cloudflare's previous approach was blunt: detect objects, write a caption, then embed the caption, which made images searchable only through whatever the caption happened to describe. At query time AI Search now checks whether an instance's embedding model handles images; if it does, the query image is embedded directly into the same vector space as the indexed images and text. Text-only models still accept an image query, but it is converted to text with ToMarkdown first, which gives every model basic multimodal support while models with native image handling keep the full visual signal.

What it costs

Billing for AI Search starts on 1 November 2026. Cloudflare says it will continue to offer a free tier on all Workers plans, and that search on its own blog and developer documentation runs on the same service.

Our opinion

Cloudflare's real product has always been a shorter list of things a small team has to run, and this launch is a tidy example of it. The interesting part is not the general availability but the multimodality, because captioning an image and embedding the caption was always a lossy shortcut that quietly decided what a picture was for. Embedding pixels directly, and falling back to ToMarkdown only when the model cannot see, is the more honest pipeline. The date to circle is 1 November, when the meter starts. A generous free tier is a good way to build habits, and the question now is whether search stays affordable once a corpus gets genuinely large.