We are assisting a client who analyses lots of 2 hour recordings of staff and then evaluates them. When the client does it with AI Studio, het gets great results. Exactly what he wants. When we use the API to programmatically do the same, with the same prompts etc, we get different results. Client is ready to cancel the whole project.
We are sending the same parameters and instructions to Gemini. Same model etc.
$generateData = [
'contents' => [
[
'parts' => [
['text' => $analysisPrompt],
['file_data' => ['mime_type' => $files_mime_type, 'file_uri' => $files_uri]]
]
]
],
'generationConfig' => [
'temperature' => $temperature,
'maxOutputTokens' => $max_tokens
]
];
Is there anything we are missing? Must we add topP and topK?