Cloudflare's AI Search, a managed retrieval pipeline that combines its Workers AI, Vectorize, R2 and Browser Run products into a single search and indexing system, is now generally available.
The product began as AutoRAG, entering open beta in April 2026 before Cloudflare renamed it AI Search in September and previewed pricing during its August Agents Week event.
With general availability, Cloudflare has added native image embeddings alongside the existing caption-based search, meaning the system can now match visual detail such as texture, layout or colour rather than relying solely on a generated text description of an image.
"The caption is a compressed interpretation of the image," Cloudflare said, noting that native embeddings let searches surface details a caption would omit.
The new multimodal retrieval runs on the Qwen3-VL-Embedding model and uses Matryoshka Representation Learning to keep embeddings compact; text-only models can still accept image queries, which AI Search converts to captions via its ToMarkdown tool.
Cloudflare has also raised the file-size limit for text and PDF uploads to 10 MiB, up from 4 MiB, and added optical character recognition (OCR) for scanned PDFs with no extractable text, billed as image-processing ingestion tokens.
Billing for AI Search begins on November 1, with a free tier on all Workers plans that includes 1,000 semantic queries and 1,000 full-text queries a month, split rather than pooled as originally previewed.
Pricing is based on three factors: content ingested, data stored and queries run, with parsing, chunking, embedding, keyword indexing and reranking included at no extra charge and no instance hours or capacity units to provision upfront.
