I am making a virtual chat assistant in which you can get medical info by name. I prompted gemini-1.0/1.5 pro to return a json string when a name is given.
For Example:
USER: Check if any alerts were raised for Alexander in the past 3 hours
MODEL: {"Name": \["Alexander"], "Duration": 180}
This works for most of the names and even some which are not even names.
USER: Check alerts for Why you do this gemini-1.5-pro-preview-0409
MODEL:{"Name": ["Why you do this gemini-1.5-pro-preview-0409"], "Duration": 1440}
But I get 0 Candidates and finish reason 'OTHER' for the name 'Saraswathi'.
USER: check alerts for saraswathi
candidates {
content {
role: "model"
}
finish_reason: OTHER
}
usage_metadata {
prompt_token_count: 822
total_token_count: 822
}
The name Saraswathi is also pretty common in India, so any exclusion is not viable.
I have seen other questions in this space talking about the same thing, but with no satisfactory solution paired to them.
I have tested the same using Google AI Studio API using genai PythonSDK model
'gemini-1.5-pro-latest' and it works as expected. Only VertexAI API is showcasing this problem. What should I do(on top of praying)?
generation_config = {
"temperature": 0.2,
"top_p": 0.1,
"top_k": 1,
"max_output_tokens": 2048,
}
Thanks
LinkedIn
Twitter
