Docs should point out that only the EXTRACTIVE answer style will produce a "content" field if the answerable probability is low

[roschler]

On this page:

https://ai.google.dev/api/rest/v1beta/models/generateAnswer

In the Response Body section, there is this text for the answer field:

"Candidate answer from the model.

Note: The model always attempts to provide a grounded answer, even when the answer is unlikely to be answerable from the given passages. In that case, a low-quality or ungrounded answer may be provided, along with a low answerableProbability."

While it is true that you will always get a Candidate answer object back, in the ABSTRACTIVE and VERBOSE answer styles, where the model is forced to synthesize text, you you will not get a content child object field containing answer text. When the answerableProbabillity is too low, you will only get a content child object with answer text in the EXTRACTIVE answer style, since the model just has to produce the grounding answer that is the closest match and does not have to synthesize any text to serve as a reply. So if you use the word "answer" in the context of a Candidate object, as the docs use it, then the docs are correct. But in the traditional sense of an answer containing text that replies to a question, the doc text is less helpful then it could be.

I suggest mentioning this in the docs, because I spent a fair amount of time trying to figure out why I didn't get any usable answer text due to this model behavior, until I wrote a rigorous test harness that tries the exact same user query and inline grounding passages set with each of the 3 AnswerStyle values. A short doc note about this could save other developers some confusion.

[hsubox76]

This is for the REST API, as this SDK doesn't currently have a generateAnswer method. Let me see if I can find out where to route this bug to.