Amazon Bedrock Adds Web Search for OpenAI GPT Models

AWS describes Web Search as an Amazon-built service backed by a web index and knowledge graph. The launch announcement says the index spans tens of billions of documents and is continually refreshed; the accompanying AWS technical documentation describes an index and cache operated within the Amazon Bedrock boundary.

Rather than passing complete pages to a model, the service extracts semantic snippets intended to fit efficiently within the model context window. Amazon Bedrock documentation explains that a supported model can invoke the tool when it determines current information is needed, receive observations containing a title, URL, and content snippet, and issue further searches in the same turn if necessary. The model then composes an answer using those results and adds source citations.

AWS documentation also states that, under the default policy, request data does not leave the AWS boundary for retrieval. IAM policies can enforce this boundary. That distinction is particularly relevant to teams whose existing web-grounding stack routes prompts or queries through an external search API.

Before this release, AWS notes, adding web grounding commonly required a third-party search provider, separate credentials and billing, custom orchestration, and vendor compliance review. Bedrock's managed approach removes those integration components for supported models, although teams still need to evaluate answer quality, source relevance, citation handling, tool-use behavior, and regional availability in their own workloads.

The main technical change is the shift from an application-managed retrieval loop to a provider-managed tool call. Comparable managed grounding services can reduce deployment work, but they also make observability important: production teams generally need to log searches, inspect citations, measure latency and answer accuracy, and test behavior on queries where no authoritative web result exists.

The service addresses a common limitation of foundation models: their trained knowledge may be stale or incomplete for current events, changing documentation, prices, regulations, and specialized facts. AWS's documentation states that Web Search is intended for such requests and that models can report when retrieved results do not support an answer, rather than filling the gap solely from parametric knowledge.

AWS has published technical documentation and pricing information for the feature. The company did not provide pricing figures in the launch announcement.