Amazon Bedrock has introduced a new feature that allows foundation models to access current web knowledge, helping them answer questions about recent events or data not included in their training. This capability, known as Web Search on Amazon Bedrock, is designed to ground model responses in real-time information, reducing the risk of hallucinations. The tool is available as a server-side built-in function, eliminating the need for third-party integration and external APIs. This update aims to streamline the development process and enhance the reliability of model outputs by providing up-to-date context during inference.
Web Search on Amazon Bedrock uses a combination of a continuously updated web index and a built-in knowledge graph to provide accurate answers to factual questions. The system retrieves relevant snippets from web pages and presents them to the model in a format optimized for its context window. This approach minimizes the amount of irrelevant information the model must process, improving both efficiency and accuracy. Additionally, the tool supports enterprise compliance by ensuring data remains within the AWS environment, with no data egress by default. Developers can enable Web Search with minimal configuration, using the OpenAI Responses API to integrate the feature into their applications.
The introduction of Web Search on Amazon Bedrock follows a broader trend in the industry to enhance the reliability and accuracy of AI models by incorporating real-time data. This update reflects Amazon's commitment to improving the capabilities of its machine learning services and addressing the limitations of static training data. By making web search a native feature of Amazon Bedrock, the company aims to simplify the development process and provide developers with a more robust toolset for building reliable AI applications.
Source: awsml
