Structured Outputs on Amazon Bedrock

This project is for educational purposes and it is not designed for production use. For production use, please implement proper security, scalability, and compliance measures appropriate for your use case.

Structured Outputs on Amazon Bedrock

Minimal, runnable examples of getting JSON-schema-constrained output from models on Amazon Bedrock, across both Bedrock endpoints and the mechanisms available on each.

Overview

Anthropic’s own structured outputs documentation states that output_config was, at launch, only available on Bedrock through a legacy integration, for a short list of models – not through Converse at all. That’s no longer the full picture: this snippet shows two independent mechanisms that work far more broadly today, plus the OpenAI-compatible Responses API, which supports structured output identically on both bedrock-runtime and bedrock-mantle.

Each script is a flat, single-purpose example – no classes, no CLI framework, nothing to configure beyond the constants at the top:

  • 01_converse_bedrock_native.py – Converse with Bedrock’s own schema field (outputConfig). Cross-provider: verified against Anthropic, OpenAI, and Moonshot (Kimi) models.
  • 02_converse_anthropic_native.py – Converse with Anthropic’s native output_config field, tunneled through additionalModelRequestFields. Anthropic-only.
  • 03_invoke_model_anthropic_native.py – InvokeModel with the same Anthropic-native output_config field, this time as a plain body field.
  • 04_responses_api.py – OpenAI-compatible Responses API. Same code works on either Bedrock endpoint; just swap BASE_URL and the model id format.
  • structured_output_on_bedrock.ipynb – all four, with explanation and a summary table of what’s supported where.

Tags

  • bedrock
  • structured-outputs
  • python
  • jupyter
  • json-schema

Technologies

  • Python 3.9+
  • AWS SDK (boto3)
  • Amazon Bedrock
  • OpenAI SDK (for the Responses API examples)
  • aws-bedrock-token-generator

Difficulty

Easy

Prerequisites

  • Python 3.9+
  • An AWS account with Bedrock model access, and credentials available via the standard AWS credential chain (aws configure, environment variables, etc.)

Setup

# Clone the repository
git clone https://github.com/aws-samples/sample-ai-possibilities.git
cd sample-ai-possibilities/snippets/structured-output-bedrock

# Install dependencies
pip install -r requirements.txt

# Run any individual example
python 01_converse_bedrock_native.py

# Or open the notebook
jupyter notebook structured_output_on_bedrock.ipynb

Cleanup Instructions

No resources are created – these scripts only make inference calls. Nothing to clean up.

Security

See CONTRIBUTING.md for security issue notifications.

Notes

This project is for educational purposes and it is not designed for production use. For production use, please implement proper security, scalability, and compliance measures appropriate for your use case.

View Source

View on GitHub