response.text could check the finish_reason and safety filters and give a more directly helpful error message.

[MarkDaoust]

Description of the feature request:

response.text could check the safety filters and give a more directly helpful error message.

If a message is blocked because of safety, currently you get an index error or something, but if the finish_reason is bad we could give some details on why text is not filled in.

[NickEfthymiou]

I would like to point out one use case where this feature request is critical. The first message received from the model is sometimes (frequently enough to be noticeable):

{
  "candidates": [
    {
      "finishReason": "RECITATION",
      "index": 0
    }
  ]
}

That’s all the bytes the client has to work with.

Assume a chat client for example. The client has issued one message, which has a system instruction "You are a helpful assistant" and an initial prompt "Introduce yourself with a friendly greeting". This normally has the effect: the model responds with "Hello, I am Bard (usually adds a waving hand emoji)" and a few more lines about the training the model has received. Plus, it generally concludes with "How can I help you today?"

The chat application then takes this response and populates the intro screen, which gets the chat conversation going. Except, if the first message back is what occasionally happens, a RECITATION block, there’s nothing to populate the intro screen with!

I have a workaround, I set up a default intro message (hardcoded in the client). If the first response is valid, the actual response overwrites the hardcoded string and that’s what the intro screen shows. If the first response is blocked, then nothing overwrites the hardcoded intro message and that’s what is shown.

It doesn’t seem right that client applications would have to do such contortions. If a suitable message were provided by whatever tier is responsible for the RECITATION block, which is totally beyond the control of anything the client can do, it would certainly help.

[zakcali]

additional note: after getting RECITATION error, chat sessions ends with nodejs api. Because last element of history contains empty parts as follows:

{ parts: [], role: 'model' }

and if you want to continue to chat, api always gives that error:

GoogleGenerativeAIError: [400 Bad Request] * GenerateContentRequest.contents[3].parts: contents.parts must not be empty