SOURCE

Cloudflare AI Search: give your agents a search engine for your data

Properties1
tagscloudflare, ai-search, mcp, knowledge-base, primary-source

2 min read

Cloudflare AI Search: give your agents a search engine for your data

Summary

Cloudflare AI Search packages crawling, ingestion, storage, embedding, hybrid retrieval, and reranking into a managed search layer for files and owned websites. A namespace can combine multiple search instances and expose them through a Worker binding or public /search and /mcp endpoints. Public endpoints can use a custom domain; Cloudflare Access can protect private instances. Website ingestion can now discover pages without a sitemap through Browser Run crawling, and the Cloudflare-AI-Search bot identifies itself and follows robots.txt.

Cloudflare demonstrates the system with a Dev Stack MCP that searches ten Cloudflare-owned documentation surfaces and returns cited, source-tagged chunks. The no-code public endpoint is appropriate for shareable search; a Worker is appropriate when search belongs inside an application or MCP server. AI Search uses hybrid semantic and keyword retrieval, making it suitable for both conceptual questions and exact named lookups.

At publication AI Search was usable but still beta/free, with preview pricing subject to change. The proposed model included free default embedding/reranking, ingestion and storage charges, and per-query charges; Cloudflare’s worked 20,000-document/30,000-query example was roughly $35 for the initial month and about $21 for later query-heavy months after free allowances. For this business, AI Search lowers the build cost of Archive-to-Agent-Tools and client knowledge-base prototypes, but search alone is not a product. The sellable layer remains one painful workflow, curated source quality, citations, permissions, and a result such as a critique, answer, or approved action. Public unauthenticated endpoints must not be used for private client material.

Key Claims

  • AI Search supports multi-instance namespaces and /search plus /mcp endpoints.
  • Custom domains and Cloudflare Access can brand or protect endpoints.
  • Hybrid search combines semantic and exact-keyword retrieval.
  • Website discovery can operate without a sitemap and respects bot policy.
  • Availability and pricing were beta/preview and subject to change.

Connections

Contradictions

  • Reinforces the existing warning that an undifferentiated vector/search box is infrastructure, not a complete archive-to-agent product.