weiqingy commented on PR #1215: URL: https://github.com/apache/flink-agents/pull/1215#issuecomment-6137521595
@wenjin272 Thanks for the review! I merged main, which brings in the `ChatResult` refactor from #1185, and added the e2e tests in 63d8c709. They are `ReActAgentTest#testReActAgentNativeStructuredOutput` and `test_react_agent_native_structured_output_on_remote_runner`. Each test runs the ReAct tool loop on Ollama (`qwen3:1.7b`) with a POJO / `BaseModel` schema under `AUTO`. A recording connection captures every request, and the tests check three things: - the model called a tool, and the loop requests carried the tools and no schema; - the last request carried the schema and no tools; - the output was parsed into the schema type with the right value. The prompt path never sends a schema, so a fallback run fails. I checked this by forcing the `PROMPT` strategy, and both tests failed. The Java test checks the schema handed to the connection's `chat`, not the wire request, because the Ollama request building isn't public. `OllamaChatModelConnectionTest#buildRequestSetsFormatForPojoSchema` already covers the POJO → `format` step, so I'd keep it as is. WDYT? Like the other ReAct e2e tests, the new ones rely on the small model calling the tool and getting the math right, so they can need a rerun. The Python one needed two on flink-2.2. -- 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]
