Switching from Claude to Nova

Akira Sato — Mon, 28 Apr 2025 11:25:29 +0000

What do I want to do?

I was using the features of the following repository to notify Slack about AWS updates. However, since this application was running on Claude, I tried to see if it could be changed to Nova.

Why Nova

Nova has multiple models for various purposes, and above all, it has a significant cost advantage.

see：https://aws.amazon.com/bedrock/pricing/?nc1=h_ls

Let's try it

The article summary is created on line 124 of hogehoge.py in the repository.

def summarize_blog(
    blog_body,
    language,
    persona,
):
    """Summarize the content of a blog post
    Args:
        blog_body (str): The content of the blog post to be summarized
        language (str): The language for the summary
        persona (str): The persona to use for the summary

    Returns:
        str: The summarized text
    """

    boto3_bedrock = get_bedrock_client(
        assumed_role=os.environ.get("BEDROCK_ASSUME_ROLE", None),
        region=MODEL_REGION,
    )
    beginning_word = "<output>"
    prompt_data = f"""
<input>{blog_body}</input>
<persona>You are a professional {persona}. </persona>
<instruction>Describe a new update in <input></input> tags in bullet points to describe "What is the new feature", "Who is this update good for". description shall be output in <thinking></thinking> tags and each thinking sentence must start with the bullet point "- " and end with "\n". Make final summary as per <summaryRule></summaryRule> tags. Try to shorten output for easy reading. You are not allowed to utilize any information except in the input. output format shall be in accordance with <outputFormat></outputFormat> tags.</instruction>
<outputLanguage>In {language}.</outputLanguage>
<summaryRule>The final summary must consists of 1 or 2 sentences. Output format is defined in <outputFormat></outputFormat> tags.</summaryRule>
<outputFormat><thinking>(bullet points of the input)</thinking><summary>(final summary)</summary></outputFormat>
Follow the instruction.
"""

    max_tokens = 4096

    user_message = {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": prompt_data,
            }
        ],
    }

    assistant_message = {
        "role": "assistant",
        "content": [{"type": "text", "text": f"{beginning_word}"}],
    }

    messages = [user_message, assistant_message]

    body = json.dumps(
        {
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": max_tokens,
            "messages": messages,
            "temperature": 0.5,
            "top_p": 1,
            "top_k": 250,
        }
    )

    accept = "application/json"
    contentType = "application/json"
    outputText = "\n"

    try:
        response = boto3_bedrock.invoke_model(
            body=body, modelId=MODEL_ID, accept=accept, contentType=contentType
        )
        response_body = json.loads(response.get("body").read().decode())
        outputText = beginning_word + response_body.get("content")[0]["text"]
        print(outputText)
        # extract contant inside <summary> tag
        summary = re.findall(r"<summary>([\s\S]*?)</summary>", outputText)[0]
        detail = re.findall(r"<thinking>([\s\S]*?)</thinking>", outputText)[0]
    except ClientError as error:
        if error.response["Error"]["Code"] == "AccessDeniedException":
            print(
                f"\x1b[41m{error.response['Error']['Message']}\
            \nTo troubeshoot this issue please refer to the following resources.\ \nhttps://docs.aws.amazon.com/IAM/latest/UserGuide/troubleshoot_access-denied.html\
            \nhttps://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html\x1b[0m\n"
            )
        else:
            raise error

    return summary, detail

By default, it uses InvokeAPI, so it's not easy to switch between multiple models. Therefore, I tried to switch the model to Nova and rewrote it as follows.

def summarize_blog(
    blog_body,
    language,
    persona,
):
    """Summarize the content of a blog post
    Args:
        blog_body (str): The content of the blog post to be summarized
        language (str): The language for the summary
        persona (str): The persona to use for the summary

    Returns:
        str: The summarized text
    """

    boto3_bedrock = boto3.client("bedrock-runtime", region_name=MODEL_REGION)

    beginning_word = "<output>"

    persona_data = f"""
    <persona> {persona} </persona>
    """
    system = [{ "text" : persona_data}]

    prompt_data = f"""
    <input>{blog_body}</input>
    <instruction>Describe a new update in <input></input> tags in bullet points to describe "What is described", "Who is this update good for" in a way that a new engineer can follow. 
    Description shall be output in <thinking></thinking> tags and each thinking sentence must start with the bullet point "- " . 
    Make final summary as per <summaryRule></summaryRule> tags. 
    Try to shorten output for easy reading. 
    You are not allowed to utilize any information except in the input. 
    Output format shall be in accordance with <outputFormat></outputFormat> tags.</instruction>
    <outputLanguage> {language} </outputLanguage>
    <summaryRule>The final summary must consist of at least three sentences, including specific use cases in which it is useful.
    Output format is defined in <outputFormat></outputFormat> tags.</summaryRule>
    <outputFormat><thinking>(bullet points of the input)</thinking><summary>(final summary)</summary></outputFormat>
    Follow the instruction.
    """

    messages = [
        {
            "role": "user", 
            "content": [{"text": prompt_data}]
        }
    ]

    inf_params = {
        "maxTokens": 4096,
        "topP": 0.1,
        "temperature": 0.5
    }

    additionalModelRequestFields = {
        "inferenceConfig": {
            "topK": 20
        }
    }

    outputText = "\n"

    try:
        response = boto3_bedrock.converse(
            modelId=MODEL_ID,
            messages=messages,
            system=system,
            inferenceConfig=inf_params,
            additionalModelRequestFields=additionalModelRequestFields
        )
        # response_body = json.loads(response.get("body").read().decode())
        outputText = beginning_word + response['output']['message']['content'][0]['text']
        print(outputText)
        # extract contant inside <summary> tag
        summary = re.findall(r"<summary>([\s\S]*?)</summary>", outputText)[0]
        detail = re.findall(r"<thinking>([\s\S]*?)</thinking>", outputText)[0]
    except ClientError as error:
        if error.response["Error"]["Code"] == "AccessDeniedException":
            print(
                f"\x1b[41m{error.response['Error']['Message']}\
            \nTo troubeshoot this issue please refer to the following resources.\ \nhttps://docs.aws.amazon.com/IAM/latest/UserGuide/troubleshoot_access-denied.html\
            \nhttps://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html\x1b[0m\n"
            )
        else:
            raise error

    return summary, detail

The following is a sample executed on Nova.

Conclusion

I was able to easily switch from Claude to Nova. You can also try it for a lower cost.
I also loved how easy it was to switch models using the Converse API.
Try it out and see for yourself!

Extra