Cloudflare AI Search Goes GA With Native Image Embeddings and OCR

Cloudflare Blog · rss · 2026-10-01

Cloudflare's AI Search — a fully managed index and retrieval pipeline combining Workers AI, Vectorize, R2, and Browser Rendering — is now generally available. Key updates:

Native multimodal retrieval: instead of the old caption-only approach (detect objects → caption → embed text), AI Search now embeds image pixels directly for visual retrieval while keeping captions for text understanding. It uses the Qwen3-VL-Embedding model, so query images land in the same vector space as indexed images and text — enabling search by texture, composition, and spatial relationships. Text-only embedding models fall back to captioning query images via ToMarkdown. Matryoshka Representation Learning (MRL) keeps smaller embeddings useful.
Bigger files and OCR: file size limit raised from 4 MiB to 10 MiB; PDFs can go through OCR so scanned documents get extracted and indexed.
Billing: starts November 1, 2026, with a free tier on all Workers plans. Ingestion $0.75/1M tokens (+$0.50 for image processing), storage $2.00/GB-month, semantic queries $0.75/1k, full-text $0.10/1k. Free allotment: 5M ingestion tokens, 10GB storage, 1,000 semantic + 1,000 full-text queries. Embedding and reranking are free with select Workers AI models.

Roadmap: full video/audio ingestion, a more scalable keyword engine, and simpler indexing for Cloudflare-hosted sites so AI agents can discover and consume content more easily.

Original post →

Cloudflare AI Search Goes GA With Native Image Embeddings and OCR

More from Apps