Google gemini errors
I have a few agent and prompt steps in my workflow running Gemini 2.5 Pro but this morning it's failing. The logs show this but I don't see anything too helpful. I can't see any reports of Gemini going down but once I switched all my agent steps to Sonnet, everything worked fine again. Anyone else see this?
{
"time": 1752657554517,
"type": "debug",
"payload": {
"ref": {
"type": "agent",
"nodeID": "6845fd895dc2e341c177ca0a",
"agentID": "6845fd8b01606e704f0ef475",
"diagramID": "64dbb6696a8fab0013dba194"
},
"level": "warn",
"message": "execution failed",
"context": "Agent",
"metadata": {
"error": {
"name": "ClientError",
"stackTrace": {
"code": 404,
"status": "NOT_FOUND"
}
},
"latency": 136
}
}
}
2 Replies
Heads-up: the 404 NOT_FOUND that’s showing in the Voiceflow log is coming straight from Google’s Vertex AI / Generative AI endpoint. In practice it means “the model ID you asked for doesn’t exist (anymore) in the region tied to your Google project.”
Why it suddenly started this morning
• Google occasionally deprecates preview model IDs (e.g. gemini-1.0-pro-preview, gemini-pro-vision-preview, etc.) and replaces them with a new alias such as gemini-1.5-pro-latest.
• When the old ID disappears, any request sent with that name returns a 404, which Voiceflow surfaces as a ClientError.
• Switching your steps to Sonnet works because the Anthropic endpoint is unaffected. What to check/fix in your Voiceflow Agent or Prompt step
1. Open the AI Provider settings for each step that’s using Gemini.
2. In the “Model” dropdown make sure you’re using one of the currently valid IDs:
• gemini-1.5-pro-latest (text)
• gemini-1.5-flash-latest (cheaper, faster)
• gemini-1.0-pro-vision-latest (vision)
3. Re-run the flow. If the model ID is valid, the 404 disappears. If you need to keep older IDs working, create a quick fallback:
• Add a second Provider path (e.g. Sonnet or GPT-4o) in the same Agent step.
• Set “Error fallback” to point to that provider.
The step will automatically retry on the alternate model if Gemini throws another 404/5xx. Finally, double-check your Google project:
– The API key / service account you supplied to Voiceflow must have vertexai.generativelanguage.user role.
– The project’s region should match the model region (us-central1 and europe-west4 are the defaults). Once the model name is updated your Gemini steps should run normally again.
• Google occasionally deprecates preview model IDs (e.g. gemini-1.0-pro-preview, gemini-pro-vision-preview, etc.) and replaces them with a new alias such as gemini-1.5-pro-latest.
• When the old ID disappears, any request sent with that name returns a 404, which Voiceflow surfaces as a ClientError.
• Switching your steps to Sonnet works because the Anthropic endpoint is unaffected. What to check/fix in your Voiceflow Agent or Prompt step
1. Open the AI Provider settings for each step that’s using Gemini.
2. In the “Model” dropdown make sure you’re using one of the currently valid IDs:
• gemini-1.5-pro-latest (text)
• gemini-1.5-flash-latest (cheaper, faster)
• gemini-1.0-pro-vision-latest (vision)
3. Re-run the flow. If the model ID is valid, the 404 disappears. If you need to keep older IDs working, create a quick fallback:
• Add a second Provider path (e.g. Sonnet or GPT-4o) in the same Agent step.
• Set “Error fallback” to point to that provider.
The step will automatically retry on the alternate model if Gemini throws another 404/5xx. Finally, double-check your Google project:
– The API key / service account you supplied to Voiceflow must have vertexai.generativelanguage.user role.
– The project’s region should match the model region (us-central1 and europe-west4 are the defaults). Once the model name is updated your Gemini steps should run normally again.
Any non Tico answers?
In the Agent step settings, I only see Gemini 2.5 Pro, not the specific model. Surely this should always default to a working model.