Support for stop_reason in Bedrock (Anthropic) Responses

[cis-hotta]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Currently, Bedrock (Anthropic) responses lack the stop_reason attribute found in Anthropic responses. This makes it difficult to determine if the response was cut off due to reaching the max_token limit, leading to truncated responses.

Motivation

When using models like Claude3, responses often exceed the token limit. Adding the stop_reason attribute would allow for better detection and handling of these cases, improving error management and user experience.

Reference:
For more details, refer to the response metadata documentation.

Proposal (If applicable)

Include the stop_reason attribute in Bedrock (Anthropic) responses to indicate if the response was stopped due to reaching the max_token limit.

[cis-hotta]

Great news! The stop_reason attribute has been added to Bedrock (Anthropic) responses, allowing for better handling of cases such as token limit exceeding. This is a significant improvement in usability. Thank you for the update!

Here's an example of how it now works:

from langchain_aws import ChatBedrock
llm = ChatBedrock(model_id='anthropic.claude-3-haiku-20240307-v1:0')
llm.invoke('Hi!').response_metadata

The output now includes the stop_reason:
{
'usage': {'prompt_tokens': 9, 'completion_tokens': 12, 'total_tokens': 21},
'stop_reason': 'end_turn',
'model_id': 'anthropic.claude-3-haiku-20240307-v1:0'
}
This addition allows for easier handling of token limit exceedances and other scenarios.
One suggestion: It appears that the documentation at Response metadata | 🦜️🔗 LangChain is outdated. Updating it would help users discover this new feature more easily.