Cloudflare AI Search: Multimodal & GA

AI Search: Beyond Text

Cloudflare's AI Search moving to General Availability marks a significant step towards making powerful AI-driven search accessible to a broader developer base. The integration of multimodal capabilities, including native image embeddings and OCR for PDFs, is particularly noteworthy. This moves beyond traditional text-based search, enabling richer and more nuanced retrieval based on visual cues and scanned document content. The Matryoshka Representation Learning (MRL) approach to embeddings is a smart technical choice, balancing detail preservation with efficiency. Furthermore, the transparent pricing model, based on ingestion, storage, and queries, with a generous free tier, lowers the barrier to entry and predictability for developers, a crucial factor for adoption. The inclusion of OCR for scanned PDFs directly addresses a common pain point in document search, making previously unsearchable content accessible.

However, while the GA announcement is strong, the billing start date of November 1, 2026, feels quite distant, potentially creating a prolonged period of uncertainty for businesses planning long-term integrations. The initial pricing, though detailed, will require close monitoring to understand its real-world cost-effectiveness compared to other managed vector database or search solutions, especially as usage scales. The mention of refactoring the keyword search engine for better scalability suggests that current limitations might exist for extremely large datasets, which could be a concern for enterprise-level deployments. The focus on image and PDF multimodal support is excellent, but the roadmap hinting at video and audio processing in the future indicates that current support, while advanced, is still evolving. Developers will need to weigh the immediate benefits against the ongoing development and potential future feature sets when choosing a platform.

Key Points

• Cloudflare AI Search is now generally available, integrating Workers AI, Vectorize, and R2 into a managed pipeline.
• Key new features include native image embeddings for multimodal search and OCR for scanned PDFs, significantly expanding search capabilities beyond text.
• Matryoshka Representation Learning (MRL) is employed for efficient and detailed multimodal embeddings.
• File size limits for text files and PDFs have increased to 10 MiB.
• Billing will commence on November 1, 2026, with a clear pricing model based on ingestion, storage, and queries, alongside a generous free tier.
• Future roadmap includes support for video and audio processing, and improvements to the keyword search engine for larger datasets.

---

📖 Source: AI Search is now generally available