Cloudflare AI Search Reaches General Availability With Multimodal Retrieval

Summary

Cloudflare has made AI Search generally available as a managed indexing and retrieval pipeline built from Workers AI, Vectorize, R2 and Browser Run. The service now supports native image embeddings alongside captions, allowing visual characteristics to remain searchable instead of relying only on generated text descriptions. Cloudflare says Matryoshka Representation Learning helps keep these richer embeddings efficient, and native multimodal retrieval is available with the Qwen3-VL-Embedding model. When an account uses a text-only embedding model, image queries can still be converted to text with ToMarkdown and searched through the resulting caption. Queries can be optionally rewritten, run through vector and keyword search in parallel, fused and optionally reranked before returning chunks or passing them to a generation model. The update also raises the supported file limit for text files and PDFs from 4 MiB to 10 MiB and adds OCR for scanned PDFs, with OCR billed as image-processing ingestion. Billing begins November 1, 2026, while all Workers plans receive a free monthly allowance of 5 million ingestion tokens, 10 GB of stored data, 1,000 semantic queries and 1,000 full-text queries. Paid usage covers ingestion, storage and queries, while parsing, chunking, embedding with selected Workers AI models, indexing and reranking are included; third-party model charges remain separate. Cloudflare says it is working toward video and audio ingestion, more scalable keyword search and simpler indexing for Cloudflare-hosted websites.