Cloudflare has renamed its AutoRAG product to AI Search and introduced a set of features aimed at developers building retrieval into AI agents, according to a post published on the Cloudflare Blog. The post describes AI Search as a search primitive that bundles a vector index, keyword index, and indexing pipeline behind a single API, and says it can be invoked from a Cloudflare Worker, the Agents SDK, or the Wrangler CLI.
What Cloudflare says is shipping
The blog post lists two headline capabilities. The first is hybrid search, which Cloudflare says runs vector search and BM25 keyword matching in parallel and fuses the results in a single query. The company states that the search feature on its own blog is now powered by AI Search.
The second is built-in storage and indexing. According to the post, new AI Search instances ship with their own storage and vector index, and files uploaded directly to an instance via API are indexed without requiring a separate R2 bucket or external data source to be connected first. Cloudflare says the underlying storage and index are powered by R2 and Vectorize, its existing developer platform products.
The post also introduces a new ai_search_namespaces binding that Cloudflare says allows Workers to create and delete AI Search instances at runtime, which the company frames as a way to provision one instance per agent, per customer, or per language without redeployment. The post adds that developers can attach metadata to documents, use that metadata to boost rankings at query time, and query across multiple instances in a single call.
A support-agent walkthrough
Cloudflare illustrates the product with a worked example of a customer support agent built on the Agents SDK. In the example, a shared AI Search instance named product-knowledge is created on the Cloudflare Dashboard within a support namespace and sources its documents from an R2 bucket called product-doc, according to the post. The post describes this as the shared knowledge base every agent can reference.
Alongside the shared instance, the walkthrough shows the agent creating a separate AI Search instance per customer at runtime using the namespace binding. Cloudflare writes that each per-customer instance "starts empty and accumulates context over time" as the agent saves summaries of resolved issues, so that prior fixes are searchable on the customer's next interaction.
New instances come with their own storage and vector index. Upload files directly to an instance via API and they're indexed.
The sample agent code in the post extends AIChatAgent from the Agents SDK and defines two tools: search_knowledge_base, which queries the product docs and the customer's per-customer instance in a single call, and a save_resolution tool for writing back summaries. The example uses the model identifier @cf/moonshotai/kimi-k2.5, which the post refers to as Kimi K2.5, served via Workers AI.
Ranking controls and multi-instance queries
The post includes a code sample in which the search call passes ai_search_options containing a boost_by parameter set to order results by a timestamp field in descending order, and an instance_ids array listing both product-knowledge and the per-customer instance. Cloudflare describes this as searching across both instances at once and surfacing recent documents over older ones.
The company positions these controls as replacements for the fusion logic and per-agent index plumbing that developers would otherwise build themselves. The post states that without a managed primitive, teams setting up search for agents need "a vector index, an indexing pipeline that parses and chunks your documents, and something to keep the index up to date," plus a separate keyword index and fusion layer if keyword search is required.
Context: from AutoRAG to AI Search
AI Search is the successor to AutoRAG, the retrieval-augmented-generation product Cloudflare introduced in an earlier post titled "Introducing AutoRAG on Cloudflare," which the current announcement links to. The new post presents AI Search as a rename and expansion of that product rather than a separate offering, and directs developers to the Cloudflare Developers documentation at developers.cloudflare.com/ai-search/ for reference material.
The blog post does not enumerate pricing for AI Search, usage limits for the ai_search_namespaces binding, or benchmark figures for hybrid-search quality or latency. It also does not state whether any component of AI Search is open source; the described APIs, bindings, and dashboard tooling are commercial Cloudflare services tied to Workers, R2, Vectorize, and Workers AI.
DeepBrief contacted Cloudflare for clarification on pricing, rate limits, and independent benchmarks for the hybrid-search feature, and had not received a response at the time of publication.