AWS Launches Web Search on Amazon Bedrock for Native Foundation Model Grounding

What happened

Amazon Web Services announced the general availability of Web Search on Amazon Bedrock, a native server-side tool designed to ground foundation model responses in current web knowledge. The launch follows AWS's earlier announcement of Web Search on AgentCore at the AWS New York Summit 2026, and extends the capability directly into Bedrock's model inference layer.

The core problem AWS is addressing: foundation models can't answer questions about events, data, or changes that occurred after their training cutoff. Grounding models in live web data reduces hallucinations and closes that knowledge gap, but historically required developers to integrate and maintain third-party search vendors — adding latency to project timelines, data residency risk, and vendor security review overhead.

How it works

Web Search on Bedrock is backed by what AWS describes as a web index spanning billions of documents, refreshed continually, combined with a built-in knowledge graph that maps entities and their relationships. According to AWS, factual queries — such as who wrote a specific book or when an event occurred — are resolved against the knowledge graph rather than left to the model to infer from raw page text, which the company claims reduces small factual errors common in retrieval-augmented generation.

The tool performs what AWS calls "semantic snippet extraction," pulling only the relevant passages from web pages rather than passing full raw pages into the model's context window. This is intended to reduce token consumption on boilerplate content and keep retrieval latency low.

Enablement requires no external API keys, SDKs, or orchestration layers. Developers add a single parameter to an existing OpenAI-compatible API call:

tools=[{"type": "web_search", "external_web_access": False}]

Bedrock handles the full search lifecycle server-side: identifying when a query needs current information, formulating the search query, retrieving from the index and knowledge graph, injecting results into context, and returning a grounded response with structured citation annotations (URL and page title per source).

Access and permissions

At launch, Web Search is available only for OpenAI models served through Amazon Bedrock's next-generation inference engine, accessed via the bedrock-mantle endpoint. Authentication uses existing AWS credentials — no separate API keys — via a short-lived bearer token (up to 12 hours) derived from IAM credentials through SigV4.

Two permission sets are required: standard Bedrock inference permissions, and Web Search tool permissions (bedrock-websearch:InvokeSearch at minimum, with optional InvokeFetch for full-page reads and ExternalWebAccess for live-web retrieval). Currently, only indexed-web retrieval is served in production; live-web retrieval is planned for a future update, though the external_web_access parameter already exists in the API.

AWS states that Web Search offers zero data egress by default — customer data does not leave the AWS environment — positioning the tool for compliance-sensitive enterprise use cases. The company notes future capabilities may expose data only at explicit customer request.

What this means

This is AWS folding a capability — real-time grounding — that developers previously had to source from third parties (Google, Bing, Brave, Tavily, Perplexity's API, etc.) directly into Bedrock's inference stack. The pitch is reduced integration overhead and a cleaner compliance story, since data doesn't cross into a separate vendor's infrastructure.

The catch: at launch this only works with OpenAI models running through Bedrock's newer inference engine, not Anthropic's Claude or other Bedrock-hosted models, and live-web retrieval isn't yet enabled — only Amazon's pre-indexed corpus is live. That limits immediate utility for teams using Bedrock's broader model catalog. Pricing for the Web Search tool itself was not disclosed in AWS's announcement. Expect competitors — Anthropic, OpenAI, and Google — to keep pushing their own native web-grounding tools, making this a contested feature category rather than a unique differentiator for AWS.