Fix Amazon Nova support (use StrOutputParser)

[dlqqq]

Description

• Closes Fix Amazon Nova support #1198.
• Appends StrOutputParser to the runnable defined in DefaultChatHandler.
  • The original issue was caused by the runnable yielding AIMessageChunk objects with a unique structure that we did not expect. This only happens when Amazon Nova is used through ChatBedrock, as far I'm aware.
  • By adding StrOutputParser to the runnable, the runnable now yields primitive strings when streamed, which we already handle in BaseChatHandler.stream_reply().

Demo

Screen.Recording.2025-01-15.at.3.20.46.PM.mov

Open questions

(mainly for @3coins)

1. Do you know why this change fixes the original issue?
2. Are you aware of any risk this change introduces?
   • I've verified that these changes do not break support for OpenAI & Anthropic models.

[dlqqq]

Interestingly, all of the other slash commands work with Amazon Nova, even though this PR only changes DefaultChatHandler. I believe this is because the original issue in #1198 only occurs when streaming, and none of the slash commands support streaming yet.

The only exception is /fix in Jupyter AI v2. I've already made sure to update /fix in the 2.x port (#1203).

[srdas]

The code changes look good. The use of StrOutputParser() works great!

Tested for use in magics:

Note the interesting new layout of the response from StrOutputParser().

Tested for use in Chat:

Retested with Nova-Micro:

Restested with Nova-Lite:

[3coins]

@dlqqq
LGTM!
The original code expecting chunk.content to be a string might not cover all Message responses, which is causing this error.
Using StrOutputParser should handle these different content types and any future changes that might happen in LangChain.

[dlqqq]

Thanks for the reviews folks! Merging.