About Model Selection

Gemini

Thanks to the generous Google Cloud Platform, which provides good service to developers.

See more at: https://ai.google.dev/pricing

Please note that for non-professional users, it’s recommended to use Google’s Gemini model rather than Vertex AI.

Vertex AI is a complex professional platform. For local execution, you even need to install gcloud, and to use it on Dify, you need a Service Account Key in base64 format.

If you really want to try it and aren’t afraid of errors or can manually resolve issues, then you’re welcome to experiment!

Google will provide you with a json file, but Dify requires input in base64 format.

I previously created a simple Python script to execute this conversion. Brave souls are welcome to use this stepping stone I’ve created to boldly try. tools\key1ToBase64.py

Llama 3.1

Llama is a top-tier open-source model from Meta. When you choose it, you might get some strange content.

• This is not Dify’s problem; content generation is determined by the model.
• It’s also not Llama’s problem; this is a feature.

Llama is like a baby, it knows nothing, so it’s best to:

• Specify the output language
• Provide clear requirements
• Provide examples
• Or use COT - Chain of Thought

In other words, Llama 3.1 is indeed very powerful, but it almost certainly needs prompt engineering.

In fact, in the paper published by Meta about Llama 3.1’s effects and scores, the Llama team used techniques like few-shot learning and COT.

In contrast, the currently best-performing closed-source model series is Claude, which can be said to have used prompts in advance, so of course the results are good. (personal understanding)

“We tried to give Claude traits” Claude’s Character \ Anthropic