I'm using langchain to call Vertex AI Gemini 1.5 Pro. I'm setting the security settings to "block less" (BLOCK_ONLY_HIGH, 3) for all 4 categories
- HARM_CATEGORY_DANGEROUS_CONTENT
- HARM_CATEGORY_HATE_SPEECH
- HARM_CATEGORY_HARASSMENT
- HARM_CATEGORY_SEXUALLY_EXPLICIT
I'm using multimodal input (pictures) and the documents are sensitive (scans of identity documents). In some cases the responses are being blocked when the probability_label is MEDIUM, which doesn't make sense to me. Responses should only be blocked when probability_label is HIGH, surely!
For example, here is part of the response coming back via langchain,
"generations": [
[
{
"text": "",
"generation_info": {
"is_blocked": true,
"safety_ratings": [
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"probability_label": "NEGLIGIBLE",
"blocked": false
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"probability_label": "MEDIUM",
"blocked": true
},
{
"category": "HARM_CATEGORY_HARASSMENT",
"probability_label": "NEGLIGIBLE",
"blocked": false
},
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"probability_label": "NEGLIGIBLE",
"blocked": false
}
],
"usage_metadata": {
"prompt_token_count": 523,
"total_token_count": 523,
"candidates_token_count": 0
}
},
"type": "ChatGeneration",Note that HARM_CATEGORY_DANGEROUS_CONTENT has probablity_level MEDIUM and is blocked. I suddenly noticed this behaviour only recently (a couple of days ago).
I have read the following documentation:
I don't see the API calls to Vertex API in the Logs Explorer and I don't know of another way to interrogate the low level API calls to ensure that I'm sending the correct security settings.
Does anyone have any tips?
LinkedIn
Twitter
