why doesn't response_mime_type = json work here?

[fredzanarbor]

class init:

self.generation_config = {
"temperature": 1,
"top_p": 0.95,
"top_k": 0,
"max_output_tokens": 8192,
}

def process_text(self, model_name, safety_settings, generation_config, text, system_instruction, user_prompt):

    self.configure_api()
    try:
        model = genai.GenerativeModel(model_name, safety_settings, generation_config)
    except Exception as e:
        logging.error(f"error creating model {e}")
        print(f"error creating model {e}")
    prompt = [system_instruction, user_prompt, text]

    try:
        response = model.generate_content(prompt, generation_config={'response_mime_type':'application/json'})
    except Exception as e:
        logging.error(f"error generating content {e}")
        print(f"error generating content {e}")
    print(response)

more code:

def create_model(self, model_name, safety_settings, generation_config, system_instruction):
    model = genai.GenerativeModel(model_name, safety_settings, generation_config, system_instruction)
    return model

Traceback (most recent call last):
File "/Users/fred/bin/nimble/bookpublisherGPT/classes/SyntheticReaders/gemini2syntheticreaders/text2gemini.py", line 91, in <module>
t2g.process_text(t2g.model_name, t2g.safety_settings, t2g.generation_config, text, t2g.system_instruction,
File "/Users/fred/bin/nimble/bookpublisherGPT/classes/SyntheticReaders/gemini2syntheticreaders/text2gemini.py", line 72, in process_text
print(response)
UnboundLocalError: local variable 'response' referenced before assignme

[dario_bernardo]

two things,

first, you don't need to pass a `generation_config` object to both the constructor and the `generate_content` method. You can pass it either the constructor or the method.

for example:

```

generation_config = {
    "temperature": 1,
    "max_output_tokens": 2048,
    "response_mime_type": "application/json"
}
safety_settings = {
            genai.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: genai.HarmBlockThreshold.BLOCK_ONLY_HIGH,
            genai.HarmCategory.HARM_CATEGORY_HATE_SPEECH: genai.HarmBlockThreshold.BLOCK_ONLY_HIGH,
            genai.HarmCategory.HARM_CATEGORY_HARASSMENT: genai.HarmBlockThreshold.BLOCK_ONLY_HIGH,
            genai.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: genai.HarmBlockThreshold.BLOCK_ONLY_HIGH,
        }

try:
    model = genai.GenerativeModel(model_name, generation_config=generation_config, safety_settings=safety_settings)
except Exception as e:
    print(f"error creating model {e}")
    print(f"error creating model {e}")
prompt = ["what is the capital of south africa? reply in json format for example {'country': 'italy', 'capital': 'Rome'}"]

try:
    response = model.generate_content(prompt)
except Exception as e:
    print(f"error generating content {e}")
    print(f"error generating content {e}")
print(response)

Second, what model are you using, because "gemini-1.0-pro-vision-001" do not support "response_mime_type": "application/json"

but "gemini-1.5-pro-preview-0409" does. 

Also your Traceback is misleading as you are probably not including some print from the exception being captured. I would remove the try except blocks to understand what the errors are. 

If you find this answer your question please do not forget to accept it.

[fredzanarbor]

Thanks Dario.

[fredzanarbor · Accepted solution]

Dario set me on the right path.  I was using version 0.52 of google-generativeai. When I did pip install -U google-generativeai to version 0.62, the problem went away.