Cloudflare announced the general availability of AI Search, a managed indexing and retrieval platform that combines Workers AI, Vectorize, R2, and Browser Run. The service expands search beyond text to images and scanned documents, with billing beginning on November 1, 2026, and a monthly free allowance continuing across all Workers plans.
Search That Handles the Image Itself
AI Search now supports native embeddings of image pixels, while also retaining the image’s textual description. Cloudflare says this approach preserves details that may be lost when an image is reduced to a description, such as colors, texture, document layout, and spatial relationships within graphics and diagrams.
The service uses Matryoshka Representation Learning to maintain smaller, more storage- and search-efficient representations. Native multimodal retrieval is currently available with the Qwen3-VL-Embedding model. When using a model that supports images, the query image is embedded directly into the same vector space as the indexed images and text.
If the embedding model is text-only, AI Search converts the query image into text using ToMarkdown and then searches based on the resulting description. This keeps image queries broadly available across all models at a basic level, while multimodal models receive the full visual signal.
Larger Files and Reading Scanned Documents
Cloudflare has increased the maximum size for text files and PDF files from 4 megabytes to 10 megabytes. It is also now possible to enable optical character recognition, OCR, for PDF files consisting of scanned images that contain no extractable text; their pages are read before the content is split and embedded.
These capabilities can be used for searching internal documents, matching screenshots, indexing diagrams and graphics, discovering products, and finding images that are visually similar or combine visual characteristics with a textual description.
What Changes in Practice?
AI Search is shifting from a system based primarily on text or image descriptions to a retrieval pipeline that can combine vector search and keyword search, then merge and optionally rerank the results before returning them or passing them to a generative model to formulate an answer.
This step matters to development teams that manage repositories containing diverse types of content, because some important information does not appear in an image description or in text extracted from a document. However, native image support is tied to the embedding model being used, while users of text models rely on converting the image into a description; as a result, retrieval quality varies according to the model and the nature of the content.
Billing and Free Allowance
Charges will apply starting November 1, 2026, across three components: ingested content, stored data, and queries. The core process includes parsing, splitting, and embedding using Workers AI models, keyword indexing, and reranking, with no separate charges for instance runtime hours or monthly minimums.
- Content ingestion: $0.75 per million tokens, with a free allowance of 5 million tokens per month.
- Image processing: an additional $0.50 per million tokens.
- Storage: $2 per gigabyte-month, with 10 gigabytes free.
- Semantic search: $0.75 per 1,000 queries, with 1,000 free queries.
- Full-text search: $0.10 per 1,000 queries, with 1,000 free queries.
Cloudflare explains that embedding and reranking are free when specific Workers AI models are used, while third-party models are billed separately. The monthly free ingestion allowance includes all supported file types.
The company later plans to add full video and audio processing, improve the scalability of the keyword search engine for large data repositories, and simplify index creation for sites hosted on Cloudflare.