weiqingy commented on PR #1120:
URL: https://github.com/apache/flink-agents/pull/1120#issuecomment-5690678767

   Thanks for the review.
   
   Yes for the chat model layer, in both languages. Every connection now 
reports and applies native structured output: Java has Ollama, OpenAI, Azure 
OpenAI, vLLM, Anthropic, Bedrock, Gemini and watsonx, and Python has Ollama, 
OpenAI, Azure, vLLM, Anthropic, Tongyi and watsonx. Gemini and Bedrock exist 
only in Java, and Tongyi only in Python, so those are connection gaps rather 
than structured output gaps.
   
   Two things are still open. The agent path does not use it yet. A 
`ReActAgent` output schema reaches `ChatModelAction`, but the call goes through 
`BaseChatModelSetup.chat`, which invokes the connection without the schema 
(`BaseChatModelSetup.java:200`, and the same in `BaseChatModelSetup.chat` on 
the Python side). The schema is applied to the response afterwards, so an agent 
with an `output_schema` still gets the prompt fallback rather than provider 
enforcement. That wiring is #912.
   
   The other is user documentation. There is no structured output section in 
the docs yet, and the per-provider behavior now has things worth writing down, 
such as putting `@JsonProperty` on every constant of an enum used in a Java 
schema. I am drafting that section and will send it as a separate PR, then pick 
up #912.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to