Amazon Web Services says Amazon Bedrock now supports structured outputs for all Anthropic Claude models available through its bedrock-runtime and bedrock-mantle endpoints. The feature is intended to return JSON that conforms to a developer-defined schema when Claude is accessed through APIs including Boto3, the Bedrock runtime APIs, or the Messages API.

AWS describes two mechanisms. JSON Schema output format controls the structure of a model response, while strict tool use validates tool parameters against a defined schema. According to AWS, the service uses constrained decoding during token generation, rather than validating or correcting the result afterward. The company says this can reduce the need for custom parsing, retries, and post-processing.

The feature supports batch inference, cross-Region inference on the bedrock-runtime endpoint, and streaming structured responses as typed events, AWS says. When a new schema is used, Bedrock validates it against a supported subset of JSON Schema Draft 2020-12 and compiles a grammar artifact. AWS says the compiled grammar is cached for 24 hours per account; changing the schema structure triggers a new compilation, while changing only its name or description does not.

AWS lists data extraction, classification, contract analysis, support-ticket routing, and agent tool calls among potential applications. Its examples include extracting fields from documents, converting clinical notes into electronic health records, and ensuring tool inputs match specified types and required fields.