ChatBedrockConverse ignores temperature value with Claude 3.5

[DHall19357]

Description

When using ChatBedrockConverse with claude sonnet 3.5 I am getting different responses with very low temp, this is not occurring when i am using a direct boto3 call with the same inference parameters.

Example Code

llm = ChatBedrockConverse(
  model="anthropic.claude-3-sonnet-20240229-v1:0", 
  temperature=0.01, 
  max_tokens=512, 
  top_p= 0.5
)
llm_with_retries = RunnableRetry(
    bound=llm, max_attempts=attempts, min_seconds=1, max_seconds=10, jitter=True
)
response = llm_with_retries.invoke(prompt)
parsed_output = llm_parsing(
    llm=llm, parser=parser, response_content=response.content
)
parsed_output

System Info

System Information

OS: Linux
OS Version: langchain-ai/langchain#1 SMP Fri Feb 14 16:52:40 UTC 2025
Python Version: 3.10.16 (main, Jan 13 2025, 14:07:23) [GCC 11.4.0]

Package Information

langchain_core: 0.3.37
langchain: 0.3.19
langchain_community: 0.3.18
langsmith: 0.3.8
langchain_aws: 0.2.13
langchain_text_splitters: 0.3.6

[3coins]

@DHall19357
Thanks for reporting this issue. I tried your sample code, and investigated this with debugging on, and the payload to Bedrock seems to pass in the correct value of temperature. Here is the debug output:

(langchain-aws-py3.11) (base) % LANGCHAIN_AWS_DEBUG=1 python3 test_issue_450.py
2025-04-04 12:44:03,064 DEBUG | input message to bedrock: [{'role': 'user', 'content': [{'text': 'What is the capital of France?'}]}]
2025-04-04 12:44:03,064 DEBUG | System message to bedrock: []
2025-04-04 12:44:03,064 DEBUG | Input params: {'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'inferenceConfig': {'maxTokens': 512, 'temperature': 0.01, 'topP': 0.5}}
2025-04-04 12:44:03,064 INFO | Using Bedrock Converse API to generate response
2025-04-04 12:44:03,469 DEBUG | Response from Bedrock: {'ResponseMetadata': {'RequestId': '87a6ca02-a288-4f8e-9c21-abd64d3856af', 'HTTPStatusCode': 200, 'HTTPHeaders': {'date': 'Fri, 04 Apr 2025 19:44:03 GMT', 'content-type': 'application/json', 'content-length': '212', 'connection': 'keep-alive', 'x-amzn-requestid': '87a6ca02-a288-4f8e-9c21-abd64d3856af'}, 'RetryAttempts': 0}, 'output': {'message': {'role': 'assistant', 'content': [{'text': 'The capital of France is Paris.'}]}}, 'stopReason': 'end_turn', 'usage': {'inputTokens': 14, 'outputTokens': 10, 'totalTokens': 24}, 'metrics': {'latencyMs': 252}}

Here is the complete code sample I tested with:

from langchain_core.runnables.retry import RunnableRetry
from langchain_aws import ChatBedrockConverse


llm = ChatBedrockConverse(
  model="anthropic.claude-3-sonnet-20240229-v1:0", 
  temperature=0.01, 
  max_tokens=512, 
  top_p= 0.5
)
llm_with_retries = RunnableRetry(
    bound=llm, max_attempts=5, min_seconds=1, max_seconds=10, jitter=True
)
llm_with_retries.invoke("What is the capital of France?")

You can set the env variable LANGCHAIN_AWS_DEBUG=1 to turn debugging on for langchain_aws, which can help you see the actual payload sent to Bedrock, let us know if this helps, and report back here if you see an actual difference in payload between ChatBedrockConverse and boto3.