What does a typical API request look like for generating text using Amazon Bedrock (for instance, what parameters and payload are needed)?

A typical API request to generate text using Amazon Bedrock involves sending a structured payload to a specific model endpoint, with parameters defining the input and output behavior. The exact details depend on the foundation model you’re using (e.g., Anthropic Claude, AI21 Jurassic-2), but most follow a similar pattern. You’ll need to authenticate your request using AWS Signature Version 4 (SigV4), include headers like Content-Type: application/json, and send a JSON body containing your prompt and configuration options. For example, a request to Claude might target the anthropic.claude-v2 model endpoint with a payload specifying prompt, max_tokens_to_sample, and temperature.

The payload typically includes parameters that control text generation. A common structure includes the input text (prompt), a token limit (max_tokens or max_tokens_to_sample), and sampling settings like temperature (randomness) and top_p (diversity). For example, a Claude request might look like this:

{
 "prompt": "\n\nHuman: Write a haiku about clouds\n\nAssistant:",
 "max_tokens_to_sample": 100,
 "temperature": 0.5,
 "top_p": 0.9,
 "stop_sequences": ["\n\nHuman:"]
}

Here, stop_sequences defines text that signals the model to stop generating. Other models like Jurassic-2 might use slightly different parameters (e.g., numResults for multiple outputs or penalties for repetition control). Always check the model provider’s documentation for exact requirements.

The response is a JSON object containing the generated text and metadata. For the example above, Claude might return:

{
 "completion": "Drifting cotton wisps,\nPainting whispers on blue skies,\nSun's fleeting embrace.",
 "stop_reason": "stop_sequence",
 "stop": "\n\nHuman:"
}

Handling errors (e.g., token limits exceeded) requires checking HTTP status codes and error fields. Developers often wrap this in a function that constructs the request, handles authentication, and parses the response. Using the AWS SDK (e.g., Boto3 for Python) simplifies SigV4 signing, but you can also call the API directly with tools like curl by manually generating the required headers.