Bedrock adds schema-enforced JSON and strict tool arguments for Claude

An extraction pipeline can fail before it reaches the substance of an answer: the model returns prose where code expects an array, omits a required field or uses the wrong data type. Amazon Bedrock’s October structured-output update for Claude addresses that interface problem by constraining generation to a declared JSON schema.[1][2]

AWS announced support across the bedrock-runtime and bedrock-mantle endpoints, including strict schemas for tool arguments. Availability follows the regions where Claude models are offered. This changes the hosted output contract; it is not a new Claude model.[2]

The schema becomes a generation constraint

Bedrock validates the request against its supported subset of JSON Schema Draft 2020-12, compiles a grammar for a new structure and constrains the model’s token choices. That is stronger than a prompt asking for valid JSON: ordinary completed output is limited to the structure the application specified.[1]

Compiled grammars are cached within the customer’s account for 24 hours from first access. A new or changed structure can incur compilation latency, while the same schema reuses the cache. AWS says changes only to name or description fields leave that grammar cache intact.[1]

This favours stable interfaces. An application extracting the same fields from many documents can reuse its contract, while one generating a new schema for every request may repeatedly pay the preparation cost. The schema itself still needs to match the downstream service’s expectations.[1]

Supported types are not unlimited validation

Objects must set additionalProperties to false. The supported subset includes basic JSON types, enumerated values and certain string formats. It excludes recursive schemas, external references and numeric constraints such as minimum, maximum and multipleOf. A schema that needs those rules cannot assume the decoder enforces them.[1]

For example, an invoice field can be constrained to a number and still contain 10,000 when the source document says 1,000. A value can also have the correct type while exceeding a business limit that the schema subset does not support. Evidence checking and application rules answer those questions after formatting has succeeded.[1]

Strict tool arguments provide the same kind of boundary. A call may contain an allowed action and a well-formed customer identifier without establishing that the action is appropriate for that customer. Authorisation belongs to the service carrying out the request, rather than to the shape of the generated arguments.[1]

Clients still need an incomplete-response path

AWS identifies two cases that can return nonconforming output: a safety refusal and generation stopped at the maximum token limit. Applications must examine the stop reason before treating a response as a complete object. In a stream, receiving well-typed events likewise does not mean the final object has arrived.[1]

Report an error or suggest a correction