Support structured output in ChatDatabricks

[B-Step62]

Resolve #19

Implement with_structured_output() method in ChatDatabricks to support structured output feature compatible with OpenAI. The code mostly copies ChatOpenAI implementation with slight adjustment e.g. FMAPI doesn't support parallel_tool_calls parameter.

Test

• Unit test
• Integration test via local machine.
• Tested on Databricks:

[B-Step62]

@harupy @BenWilson2 Could you review this change that mostly copies what OAI does? Wanted to close the FR before releasing the next version by the end of this week.

[BenWilson2]

In the OAI implementation, they're doing the same thing?
IIRC their tool requests are sequential, but it will be good to double check.

[B-Step62]

Yes https://github.com/langchain-ai/langchain/blob/4fab8996cf3a5a34bd5333c6848b0bccf798a6a0/libs/partners/openai/langchain_openai/chat_models/base.py#L1234-L1236

I think the rational is that it is guaranteed to have only one tool matches with the tool_name we pass here.

[BenWilson2]

SGTM! Let's merge :)

[BenWilson2]

overall LGTM - let's verify that the tool response requests from OpenAI are always provided as sequential iterations (I BELIEVE that is how their APIs work, but let's be sure we're not inadvertently dropping multiple tool requests; for context, Anthropic adheres to a similar response structure but their ToolRequestFormat is a collection instead of sequential calls)

[B-Step62]

Reading through the code closely, the more direct reason is that LLM can only returns single tool call for structured output because we (and OAI/Anthoropic) set tool_choice parameter to only include the formatting tool🙂

https://github.com/langchain-ai/langchain/blob/4fab8996cf3a5a34bd5333c6848b0bccf798a6a0/libs/partners/openai/langchain_openai/chat_models/base.py#L1226

It's good to understand the rational behind (rather than no-brain copy paste). Great callout, Ben!