Streaming responses for Mistral models are getting truncated

[ihmaws]

Description
When using any of the Mistral models in streaming mode, the last token will get truncated if a stop sequence is added.

Reproducible Sample

from langchain_aws import ChatBedrock
from langchain_core.messages import HumanMessage, SystemMessage

messages = [
    SystemMessage(
        content="Respond in no more than 5-8 words."
    ),
    HumanMessage(
        content="Hello, how are you?"
    )
]

llm = ChatBedrock(
    model_id="mistral.mistral-small-2402-v1:0",
    streaming=True,
    model_kwargs={
        "stop": ["test"] #needed to trigger the bug; this causes Bedrock service to add last token alongside the stop_reason
    },
)

#print streamed chunks
for chunk in llm.stream(input=messages):
    print(chunk.content, end="", flush=True)

Example Truncated Response (Input = "Hello, how are you?")
I'm an AI, I don't have feelin

[ihmaws]

Thanks @3coins!

I have a backport to v0.1.18 ready (untested). I can create a PR once a branch to base off is ready: https://github.com/langchain-ai/langchain-aws/compare/v0.1.18...ihmaws:langchain-aws:dev/ihm/fix-mistral-streaming-v0.1.18?expand=1

[ihmaws]

Backport into v0.1 here: #222