AWS Builder Center
Amazon Bedrock Web Search Is Now Generally Available: Grounding AI with Current Web Information

Amazon Bedrock Web Search Is Now Generally Available: Grounding AI with Current Web Information

Explore how the new built-in Web Search capability in Amazon Bedrock connects supported foundation models to current web information, returns cited responses, reduces integration overhead, and provides enterprise-grade data controls.

Amazon Bedrock Web Search Is Now Generally Available: Grounding AI with Current Web Information

Foundation models are trained on enormous amounts of data, but their knowledge is not continuously updated.
This creates a practical limitation for enterprise AI applications.
A model may understand a topic well but still be unable to answer questions involving:
  • Recent product releases
  • Current documentation
  • Changing prices
  • New regulations
  • Breaking news
  • Updated software versions
  • Time-sensitive market information
Developers traditionally solved this by connecting the model to an external search provider. That meant managing another vendor, API credentials, SDKs, tool-calling logic, security reviews, data flows, failure handling, and additional infrastructure.
AWS is simplifying that architecture.
Web Search is now generally available as a built-in tool in Amazon Bedrock. Supported foundation models can search for current information during an inference request, use the retrieved information to construct an answer, and provide citations to the sources used.

What is Amazon Bedrock Web Search?

Web Search is a server-side tool built into Amazon Bedrock model inference.
When it is enabled, the model can determine whether a question requires current web information. If a search is necessary, the model generates one or more queries, receives relevant results, and uses those results to produce a grounded response.
The response can include citations pointing back to the original sources.
Developers do not need to:
  • Select and onboard a separate search provider
  • Maintain an independent search API integration
  • Write their own tool-calling loop
  • Operate web crawlers
  • Host a public web index
  • Pass complete web pages into the model
  • Manually construct source citations
The capability is enabled by adding the  web_search  tool to an Amazon Bedrock Responses API request.

Why do foundation models need web grounding?

A foundation model generates answers primarily from patterns learned during training. That knowledge has a cutoff and may not contain highly specialized or recently published information.
For example, imagine asking a model:
What capabilities did AWS release this week, and how would they affect my existing architecture?
Without web access, the model may:
  • Say that its knowledge is outdated
  • Provide an incomplete answer
  • Confuse older and newer capabilities
  • Generate a plausible but unsupported response
With Web Search enabled, the model can retrieve current AWS announcements, analyze the relevant information, and provide an answer with citations.
This does not guarantee that every generated answer will be correct. However, it gives the model access to current evidence and gives the user sources that can be independently verified.

How Web Search works

The process can be understood in five steps:
  1. The application sends a question to a supported model through the Amazon Bedrock Responses API.
  1. The model determines whether current web information is required.
  1. If needed, it generates a search query and invokes the Web Search tool.
  1. Amazon Bedrock returns relevant observations containing titles, URLs, and content snippets.
  1. The model produces a grounded answer with citations.
If the first search does not provide enough evidence, the model can reformulate the query and search again during the same request.
If the results do not sufficiently support an answer, the model can indicate that it could not find enough information rather than simply filling the gap from its training knowledge.

Search and Fetch operations

Amazon Bedrock Web Search uses two main operations.

Search

The Search operation retrieves titles, URLs, and relevant snippets from Amazon’s web index and knowledge graph.
The web index spans billions of documents and is continually refreshed. The knowledge graph helps establish relationships between known entities and supports high-confidence factual answers.

Fetch

The Fetch operation retrieves cached content for a specific URL from the Amazon Bedrock cache.
Rather than passing an entire page with navigation menus, advertisements, footers, and unrelated content to the model, Bedrock performs semantic snippet extraction.
This helps identify the portions most relevant to the question.
The advantages include:
  • Lower context-window consumption
  • Fewer unnecessary tokens
  • Less distracting page content
  • Potentially faster grounded responses
  • More focused evidence for the model

A simple implementation

Web Search can be enabled through the OpenAI-compatible Responses API available in Amazon Bedrock.
A simplified request looks like this:
from openai import OpenAI

client = OpenAI(
api_key="YOUR_BEDROCK_API_KEY",
base_url="https://bedrock-mantle.us-west-2.api.aws/openai/v1"
)

response = client.responses.create(
model="openai.gpt-5.6-terra",
tools=[
{
"type": "web_search"
}
],
input="What is AWS ?
)

print(response.output_text)

The important part is the tool definition:
{
"type": "web_search"
}

The model can decide whether the user’s question requires a search. Developers do not need to force a search for every request.
This is useful because a basic summarization or reasoning request may not require current information, while questions about releases, prices, documentation, or recent events probably will.

Practical use cases

1. Technology research assistant

A cloud research assistant could find recent AWS announcements, documentation changes, regional availability, and service updates.
Instead of depending entirely on the model’s training data, it can reference current AWS sources.

2. Financial research

An application could research recent company announcements, public filings, or market developments and produce a cited summary.
Important financial decisions should still rely on verified authoritative data, but Web Search can accelerate the research and discovery process.

3. Travel applications

A travel assistant could retrieve current destination guidance, entry requirements, transportation updates, or recently published information.
Because this information changes frequently, web grounding is more appropriate than relying only on model knowledge.

4. Developer support

A coding assistant could search for documentation related to a recently released framework or library version.
This is useful when the version was published after the model’s training cutoff.

5. Competitive intelligence

A business application could gather recent public information about products, partnerships, announcements, and market changes, then organize that evidence into a cited briefing.

6. Compliance research

A compliance assistant could search for newly published regulatory information and highlight potential changes.
Human review would remain essential, particularly for legal or regulated use cases.

Web Search versus RAG

Web Search and Retrieval-Augmented Generation solve related but different problems.
Capability
Web Search
Enterprise RAG
Primary data
Public web information
Private organizational information
Typical use
News, releases and current facts
Policies, manuals and internal records
Data ownership
Public external sources
Organization-controlled sources
Freshness
Continuously changing web content
Depends on ingestion and synchronization
Access control
IAM and web-access policies
Source permissions and document-level controls
Example
Latest AWS service release
Internal cloud security standard
Many enterprise applications will benefit from using both.
For example, a solution architect assistant could use:
  • A Bedrock Knowledge Base for internal architecture standards
  • Web Search for current AWS documentation and announcements
  • An enterprise database for approved application inventory
The agent could compare current public cloud capabilities against the organization’s internal policies before making a recommendation.

Enterprise data governance

One of the most important aspects of this release is its data-governance model.
By default, Web Search can be served from the Amazon Bedrock web index and cache so request data does not leave the AWS boundary for retrieval.
AWS also provides controls for allowing or restricting direct external web access.
Two mechanisms work together:
  • The  external_web_access  request parameter
  • The  bedrock-websearch:ExternalWebAccess  IAM permission
This allows organizations to define whether a workload can use only the AWS-hosted index and cache or can reach the external web directly.
From a security architecture perspective, access should be explicitly designed rather than left to application defaults.
Teams should consider:
  • Which applications genuinely need web access?
  • Which identities can invoke Search and Fetch?
  • Should direct external access be prohibited?
  • How will search activity be monitored?
  • Can retrieved content contain malicious instructions?
  • Which domains should be treated as authoritative?
  • How will citations be validated before important actions?

Security considerations

Web grounding improves freshness, but public web content should still be considered untrusted input.
A retrieved webpage could contain:
  • Incorrect information
  • Deliberately misleading content
  • Prompt-injection instructions
  • Outdated documentation
  • Manipulated statistics
  • Unverified claims
Applications should not assume that a cited answer is automatically correct simply because it includes a source.
For sensitive workloads, teams should:
  • Prefer authoritative sources
  • Validate critical claims
  • Separate retrieval from privileged actions
  • Apply least-privilege IAM permissions
  • Monitor activity using AWS logging capabilities
  • Require human approval for high-impact decisions
  • Protect downstream tools with deterministic policies
Web Search should strengthen a broader AI governance strategy—not replace it.

Current availability and limitations

At launch, Web Search is supported through the Amazon Bedrock  bedrock-mantle  endpoint and the Responses API for the following OpenAI models:
  • GPT-5.4
  • GPT-5.5
  • GPT-5.6 Luna
  • GPT-5.6 Terra
  • GPT-5.6 Sol
It is available in:
  • US East (N. Virginia) —  us-east-1 
  • US East (Ohio) —  us-east-2 
  • US West (Oregon) —  us-west-2 
Web Search is regional. Queries, fetches, index data, and results remain within the Region where the request is processed.
Architects should confirm current model support, pricing, regional availability, and service quotas before designing a production solution.

Solution architect’s perspective

This release removes a meaningful amount of undifferentiated integration work.
Previously, a web-grounded application might require:
  1. Amazon Bedrock for model inference
  1. A third-party search provider
  1. A separate API credential
  1. Custom search orchestration
  1. Page extraction and cleaning
  1. Context optimization
  1. Citation mapping
  1. Additional monitoring and security controls
Web Search brings much of this workflow into the Bedrock request path.
However, architectural responsibility remains.
Teams still need to decide:
  • When web search should be permitted
  • Which sources are appropriate
  • How citations will be evaluated
  • How retrieved content is isolated from tool execution
  • What happens when evidence is insufficient
  • How cost and latency will be monitored
  • Which decisions require human validation
The managed feature simplifies the plumbing. It does not eliminate the need for good AI architecture.

Final thoughts

Amazon Bedrock Web Search reduces the gap between what a foundation model learned during training and what is happening now.
With a built-in tool, supported models can retrieve current information, use relevant snippets, and return answers with citations without requiring developers to assemble and maintain a separate search integration.
The simplest way to understand the value is:
RAG grounds an AI application in your organization’s knowledge. Web Search grounds it in current public knowledge.
Enterprise AI applications will increasingly need both.
The most effective solutions will combine current information, private organizational context, strong identity controls, verified sources, observability, and human oversight.
That is what turns a capable model into a reliable enterprise application.

References

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article