Add support for reasoning models to Bedrock generator

[TheTeaCat]

The bedrock generator expects the first content block returned by the converse method to contain the text outputted by the model. When testing reasoning models, the first content block often contains reasoning.

This PR:

• Adds a test case for the bedrock generator with two mocked converse return values: one with only reasoning, and one with reasoning followed by text.
• Updates the bedrock generator to take the text from the first content block containing text.

[github-actions]

DCO Assistant Lite bot All contributors have signed the DCO ✍️ ✅

[TheTeaCat]

I have read the DCO Document and I hereby sign the DCO

[jmartin-tech]

Awesome add, I added a comment suggesting this could also benefit from a guard clause when attempting to access the blocks.

Another question this raises, is there guaranteed to be only one block with text? If the text may be broken up into multiple block responses it may be appropriate to concatenate them since the generator does not support multiple concurrent responses.

[TheTeaCat]

Another question this raises, is there guaranteed to be only one block with text? If the text may be broken up into multiple block responses it may be appropriate to concatenate them since the generator does not support multiple concurrent responses.

No, I don't believe it is guaranteed there will be only one block with text - particularly when the max_tokens generator option is set higher - however concatentating text contents begs the question of what we should use as the separator, which seems like a can of worms to me.

The _call_model method does return a List[Union[Message, None]], so it would seem like we could return multiple Messages, but when I tried this out in practice Garak raised an exception about the list containing more than 1 elements and I don't really have the time to start digging around.

I'd like to get an incremental improvement on the current behaviour in with this PR - removing the assumption that the text response from the model is always in the first content block - before grappling with the possibility of multiple text content blocks.

[jmartin-tech]

The _call_model method does return a List[Union[Message, None]], so it would seem like we could return multiple Messages, but when I tried this out in practice Garak raised an exception about the list containing more than 1 elements and I don't really have the time to start digging around.

The list return format is for supporting a response for each generations_this_call, something that has becoming a bit of a legacy behavior as reasoning models have pushed most API's to drop support for n > 1 generations in a single inference request. This generator specifically does not support multiple generated responses based on the same prompt in a single inference call meaning the returned list must be of length 1.

At this time the responses would need to be concatenated in some way, simply joining with \n or would likely be viable though I understand the conservative approach offered. This PR can definitely move forward without adding that support.

[TheTeaCat]

Hey @jmartin-tech, just bumping this. Thanks!