Cloudflare has made AI Search generally available, adding native image retrieval, OCR for PDFs and support for larger files. The biggest change is that image search can now work from the pixels themselves, rather than only from a generated description. There is a budget date to mark, too: Cloudflare says billing begins on 1 November 2026, with a free tier continuing across all Workers plans.
Cloudflare Watch analysis
What happened
The 1 October announcement moves Cloudflare’s managed indexing and retrieval service into general availability. The company describes uses ranging from internal-document searches to website search, and says it already uses AI Search on its own blog and developer documentation.
Previously, its image retrieval relied on detecting objects, generating a caption and embedding that text. The new approach retains captions for textual understanding while also embedding image pixels directly. That gives visual details a route into search even when a caption leaves them out.
Native multimodal retrieval is available with Qwen3-VL-Embedding. When an instance uses an image-capable embedding model, a query image is embedded into the same vector space as its indexed images and text. With a text-only model, image queries still work, but the service converts the image into a caption and searches from that description.
Cloudflare also announces PDF optical character recognition and support for larger files. Its AI Search announcement sets out the release and the forthcoming billing change.
Why it matters
A screenshot’s interface state, a product’s texture or a chart’s visual relationships may be precisely what someone needs to find. Reducing each image to a sentence can discard the distinguishing detail before the search even begins. Native image embeddings address that specific weakness.
For developers, the model choice now has a practical consequence: accepting an image query is not the same as searching its visual content directly. A text-only configuration still offers a useful fallback, but it cannot preserve details that never make it into the caption.
The billing date gives teams a concrete reason to examine their workloads before November. General availability is a useful milestone; the invoice is less interested in milestones.
Our read
This is a substantial upgrade for teams searching mixed collections of documents and images. The strongest improvement is not simply that images are supported, but that their visual information no longer has to squeeze through a caption first.
Test that distinction with images whose important differences are hard to describe: similar products with different textures, or screenshots showing different states of the same interface. Compare native retrieval with the caption fallback and check which results actually answer the task.
PDF OCR and larger-file support broaden the service’s usefulness, but the supplied announcement gives no exact file-size limits. Check those limits against your own collection before planning a migration.
What to watch
- Whether native image retrieval finds meaningful visual differences that caption-based search misses.
- How PDF OCR handles scans and layouts in real document collections.
- Which larger-file limits apply to the material teams need to index.
- How expected usage fits the free tier before billing starts on 1 November.
Discussion spark: For mixed document-and-image collections, should native visual retrieval be the default, or should teams keep caption-based search until it fails a concrete task?
Sources and evidence
- Source update (1 October 2026, 13:00 UTC)
not affiliated with or endorsed by Cloudflare