Introducing Web Search on Amazon Bedrock for foundation model grounding
AWS Anuj Jauhari ● Covered by 2 sources
AWS just made Web Search a built-in tool on Amazon Bedrock, so models can ground answers in current web info. No third-party search vendor needed, no extra API wrangling.
Based on reporting by AWS, Anuj Jauhari — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS has a habit of noticing where developers keep rebuilding the same plumbing, then absorbing that plumbing into the platform itself. Web Search on Amazon Bedrock, now generally available, is exactly that move. Instead of hunting down, integrating, and maintaining a third-party search provider just so a foundation model can answer something like a question about last week's earnings call, developers get grounding as a native Bedrock capability. It follows the earlier general availability of Web Search on AgentCore, announced at AWS New York Summit 2026, and extends the same idea into Bedrock's model inference layer.
What makes this more than a simple API wrapper is how AWS built the retrieval underneath. Web Search draws on a web index that Amazon operates and refreshes continually, spanning billions of documents, paired with a knowledge graph that maps entities and the relationships between them. For a factual question — who wrote a book, what year something happened — the tool leans on the knowledge graph for a confident answer rather than letting the model guess from scraps of page text. AWS says this cuts down on the small factual slips that creep in when an agent stitches together an answer from fragments on its own. On top of that, the system does semantic snippet extraction, pulling only the passages relevant to a query instead of dumping a full page into the model's context window, which keeps token usage down and retrieval fast.
The mechanics are refreshingly boring, in a good way. Enabling it is a single parameter added to an existing OpenAI-compatible API call — no vendor onboarding, no separate SDK, no client-side loop to build for parsing external search responses. Bedrock handles the whole lifecycle server-side: the model decides it needs current information, Bedrock formulates the search, pulls from the index and knowledge graph, injects snippets with source URLs and titles into context, and returns a grounded answer with structured citation annotations. Getting started with the OpenAI Responses API takes three steps — configure AWS credentials, point the client at the bedrock-mantle endpoint, and add the Web Search tool to the request — using existing IAM permissions rather than new API keys. At launch it's scoped to OpenAI models served through Bedrock's next-generation inference engine, with indexed-web retrieval live now and live-web fetching promised for a future update.
Compliance is where AWS spends real effort in this announcement, and it shows. By default there's zero data egress, meaning data doesn't leave a customer's AWS environment, and every call to the search and fetch permissions gets logged in AWS CloudTrail — including denied requests, with the specific condition key that caused the denial. Notably, CloudTrail doesn't record the actual query text, the URLs returned, or the page content fetched, treating queries the same as any other inference prompt. That's a deliberate split: full audit trail of who used the tool and when, without exposing what anyone was actually searching for. It's currently generally available in the US, with in-region processing in us-east-1, us-east-2, and us-west-2.
My take — AI-written commentary, not fact-checked reporting
Folding web search directly into Bedrock is a smart bet on where the friction actually lives for enterprise AI teams — not in whether grounding works, but in whether legal and security will sign off on yet another third-party vendor. AWS just made that question disappear for anyone already living inside its ecosystem, which is a bigger competitive lever than the retrieval tech itself. The catch, as always with AWS convenience plays, is that it deepens lock-in exactly as it removes friction — that's the trade every customer should go in expecting.
Read more about this at: AWS