Bedrock pixtral fix

[ajac-zero]

Title

Fix support for Pixtral Large (25.02) from Amazon Bedrock provider.

Relevant issues

Fixes #10360

Pre-Submission checklist

• [ x] I have added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• [ x] I have added a screenshot of my new test passing locally

- [ x] My PR passes all unit tests on [`make test-unit`](https://docs.litellm.ai/docs/extras/contributing_code) - [ x] My PR's scope is as isolated as possible; It only solves 1 specific problem

Type

🐛 Bug Fix

Changes

'transform_response' in 'base_invoke_transformation.py' expected Mistral models from Bedrock to produce responses with the 'outputs' key; However, Pixtral Large (25.02) responds in a different structure with the 'choices' key. The changes made aim to accommodate this variation:

1. Added a new "get_outputText" static method to "AmazonMistralConfig". This method extracts the string response from the LLM differently based on the structure of the model response. If the response holds a "choices" key, that will take priority; Otherwise, it will search for the "outputs" key. If no adequate keys are found, a Bedrock error will be raised.
2. Replaced the Mistral route for 'transform_response' in 'base_invoke_transformation.py' with this static method.
3. Added a new test file under "/bedrock/chat" to test this static method.

The reasoning behind abstracting this logic behind a static method is to facilitate unit testing in an isolated manner away from the rest of the "transform_response" method. The "AmazonMistralConfig" class was chosen to follow a similar pattern established by the use of "AmazonInvokeNovaConfig" and "AmazonAnthropicClaude3Config" within the method.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
litellm	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Apr 30, 2025 2:47pm

[CLAassistant]

All committers have signed the CLA.