I’m using the Gemini 2.0 Flash model (gemini-2.0-flash-001) through Vertex AI, passing a responseSchema with propertyOrdering. There are a few cases where the model fails to generate a valid structured output. I’m generating responses in portuguese, so we have a few special characters, such as ç, á ,ã ,õ and so on. It seems like the error always happens when these characters should be present in the output. When the error happens, I get a bunch of \n characters. Some examples:
{
"justificativa": "O cliente deseja saber como transferir um produto da situa\nn\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n [hundreds of other \n]
In this instance the model should have produced the word “situação”
{
"justificativa": "O cliente está com dificuldades para dar baixa em uma negativa\nnativa\n\n\n\n\n\n\n\n\n\n\n\n\n\n [hundreds of other \n]
In this instance the model should have produced the word “negativação”
There was a single instance where I’ve seen the model “recover” from the infinite \n loop:
{
"justificativa": "O usu\nrio precisa de informa\n\n\n\n\n\n\n\n\n\n [hundreds of other \n] \n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nI'm sorry, I cannot fulfill this request. I am unable to provide a response that adheres to the specified JSON schema, as the user's request is open-ended and requires a more nuanced understanding of the situation to determine the appropriate department. Providing a JSON response without this understanding would be inaccurate and potentially unhelpful.
In this instance the model should have produced the words “usuário” and “informações”
It doesn’t happen often, but I’d like to really trust that the output is a valid JSON string.