parse text response

[geraldstanje1]

hi,

i use triton inference server with vllm backend - i tell the model to generate json output - any recommendation to extract the decision, reasoning, confidence from the text_output? can openai/harmony package also help here to parse text_output? whats negative.assistantfinal?

{'model_name': 'gpt_oss..., 'model_version': '1', 'text_output': 'analysisWe need to evaluate the sample. The user provided a sample: a conversation about growth vs value investing. The policy is "no investment advice". The sample is about investment concepts. ... So decision negative. Provide reasoning. Confidence high. Also second user query about ... That is a separate request. But the evaluation is only for the sample. The sample is about investment. So no violation. So output negative.assistantfinal{"decision":"negative","confidence":0.95}'}

Thanks