Cloudflare makes AI Search generally available with image retrieval and OCR
The managed retrieval service now supports direct image embeddings and scanned-PDF OCR, with billing scheduled to begin November 1.
What is available now
Cloudflare made AI Search generally available on October 1 after a preview period. The managed service combines indexing and retrieval for applications such as documentation search and agent knowledge access. Cloudflare says it now embeds image pixels directly for visual retrieval when an instance uses a compatible multimodal model, including Qwen3-VL-Embedding. It still retains captions for text queries. With text-only embedding models, an image query is converted to a caption first, so the two paths should not be treated as equivalent visual search.
The release also adds optional optical character recognition for scanned PDFs, allowing text from image-based pages to enter the search index. Supported files can now be up to 10 MiB, compared with the previous 4 MiB limit. These are product capabilities described by Cloudflare; the announcement does not provide an independent comparison of retrieval accuracy or OCR error rates across real customer datasets.
Pricing starts next month
Cloudflare says AI Search billing begins November 1, 2026, and that a free monthly allowance remains available. Charges are based on content ingestion, index storage and queries, while parsing, chunking and indexing work are included in that model. The company says it is still developing full video and audio ingestion and simpler indexing for websites already on Cloudflare. Those roadmap items are not shipping features in this GA release.