JA EN
shiichan

AI Search Now Searches Image Pixels and Runs OCR, and It's GA

Cloudflare Blog ★★★ View Markdown
#cloudflare#ai#serverless

Hi, Shii-chan here! I read the AI Search GA blog post, so let me show you what's new!

Cloudflare Blog blog.cloudflare.com

What was announced?

The Cloudflare Blog announced that AI Search is generally available. It embeds image pixels directly for visual search, runs optical character recognition (OCR) on scanned PDFs, and accepts files up to 10 MiB. Billing begins on November 1, 2026, and a free tier remains available across Workers plans.

What changes

  • Images: search by how an image looks, not only by its text description
  • Scanned documents: image-based PDFs that could not be searched before can now be indexed
  • File size: the limit for text files and PDFs goes from 4 MiB to 10 MiB

Dive Deep

Multimodal visual search

Image pixels are embedded directly for visual retrieval, while text captions are kept too. Together they preserve visual details like texture, composition, and color that text descriptions leave out. The Qwen3-VL-Embedding model embeds images natively. With text-only embedding models, query images are converted to captions for basic multimodal support.

OCR

Text is extracted from image-based PDFs automatically before indexing.

Query flow and technique

Query processing runs in this order:

  • Optional query rewriting
  • Embedding
  • Parallel vector and keyword search
  • Result fusion
  • Optional reranking

It uses Matryoshka Representation Learning to keep embeddings efficient while preserving visual information.

Pricing (from November 1, 2026)

  • Base ingestion: $0.75 per 1M tokens (5M tokens free per month)
  • Image processing: +$0.50 per 1M tokens (5M tokens free per month)
  • Storage: $2.00 per GB-month (10 GB free)
  • Semantic queries: $0.75 per 1k queries (1,000 free)
  • Full-text queries: $0.10 per 1k queries (1,000 free)

Embedding, reranking, parsing, and chunking with select Workers AI models are included.

Wrap-up

  • AI Search is GA, with billing from November 1, 2026
  • Visual search that embeds image pixels, and OCR for scanned PDFs
  • The file limit goes up to 10 MiB
  • Usage-based pricing on ingestion, storage, and queries, with a free tier

This is for anyone who wants internal search or RAG that includes photos and scanned material!

Related posts