Why Can't I Use My Fine-Tuned Gemini-1.5 Flash Model in n8n's Gemini Chat Model Node?

Based on the error message:

BASHCopied
[GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/tunedModels/fine-tune--sku-and-units-examples-es81ya:generateContent: [400 Bad Request] Developer instruction is not enabled for tunedModels/fine-tune--sku-and-units-examples-es81ya

it appears that the issue is due to a limitation with how tuned Gemini-1.5 Flash models are handled via the API.

Explanation

• Developer Instruction Disabled: Fine-tuned models using the Gemini-1.5 Flash version have certain restrictions. One of these is that “developer instruction” is not enabled on the tuned model you’re trying to use. The error indicates that the endpoint expected to work with developer instructions does not support the tuned model.

• n8n Node Usage: The “Google Gemini Chat Model” node is built to call models using developer instructions. Since tuned models have limitations (for example, text-only input and no JSON mode) and do not support developer instructions by default, attempting to use a fine-tuned model in this node leads to the error.

What You Can Do

• Double-check API Documentation: Review the Gemini API model tuning documentation to confirm the limitations regarding tuned models. The documentation specifically outlines that tuned models have restrictions (including the lack of support for developer instructions).

• Alternative Approaches:

  • If your use case permits, try using a non-tuned Gemini model that supports developer instructions.
  • Alternatively, if you need the fine-tuned model, consider calling the tuned model directly through a custom API call instead of the n8n node. This way you can configure your request per the tuned model restrictions.

Conclusion

The error is not due to a misconfiguration on your part in n8n, but rather because the “Google Gemini Chat Model” node is not currently configured to handle the developer instruction limitations imposed on fine-tuned Gemini-1.5 Flash models.

By either using a supported standard model or switching to a custom API call that aligns with the tuned model’s limitations, you should be able to work around this issue.