Cloudflare makes AI Search generally available

Cloudflare announced that AI Search, its fully managed indexing and retrieval pipeline, is now generally available. The service launched more than a year ago and is built on Workers AI, Vectorize, R2 and Browser Run. Cloudflare says it uses AI Search for search on its own blog and developer docs.

The main technical change is native multimodal retrieval. The earlier version handled images by running object detection, writing a caption and embedding that text, so an image could only be found through details the caption happened to include. AI Search now embeds image pixels directly, using the Qwen3-VL-Embedding model, and still keeps captions for text matching. It uses Matryoshka Representation Learning to keep embeddings small enough for manageable storage and fast search. Users of text-only embedding models can still search with an image: the query image is converted to a caption with ToMarkdown and searched as text.

Queries can optionally be rewritten. Vector search and keyword search then run in parallel, and their results are fused and can optionally be reranked. The file size limit for text files and PDFs has risen from 4 MiB to 10 MiB. Optional OCR extracts text from scanned PDFs before chunking.

Billing starts November 1, 2026. Customers pay for ingestion, storage and queries. Parsing, chunking, embedding and reranking are included, and there are no instance hours or monthly minimums. Base ingestion costs $0.75 per million tokens, with 5 million tokens free each month. The free tier on all Workers plans now includes 1,000 semantic queries and 1,000 full-text queries, replacing the shared pool of 2,000 announced with preview pricing. Image processing costs extra. The source excerpt is cut off before the full price table, so storage and query rates are not covered here.