Create demo using gemini model

[jeremyphilemon]

Gemini Function Calling Feedback

• the model has a strict order for multi turn chat completions:
  • user → model: call function → function response → model → user
• as a result, if we choose to include the function call + response, we have to do a model generation after rendering a component, which is another recursive call.
• it would be great to handle function calls by the client without requiring another model generation, or just offer more flexibility without the order as constraints
• if we choose to remove the function call + response, the model has no memory about the interaction and messes with the sequential calls, so it requires some careful prompt injection to let the model know what has happened
• also there can't be consecutive model roles, so we can't append multiple ui interactions to the conversation history as distinct messages
• vision and text models do not follow the same messages input parameter format
• when using tools, the model does not answer to general questions
  • user: who is michael jordan?
  • model: I cannot fulfill this request. I do not have the functionality to provide information about people.
• System message support would be very useful
• Being able to force the call of a specific function would be useful (for generateObject and streamObject)
• From the docs it’s unclear if the tool call arguments are streamed with deltas
• Support for a ‘seed’ parameter would be great
• Support for JSON mode (or grammars) would be great
• Temperature parameter (ranging from 0 to infinity) is difficult to map on our temperature range 0…1
• For parallel calls, there is no clear mapping from functionCall to functionResponse (e.g. via a call id)

[vercel[bot]]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
ai-chatbot-rces	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Apr 3, 2024 6:55pm
oss-ai-chatbot	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Apr 3, 2024 6:55pm

[athrael-soju]

Really cool implementation. Are you planning to add a switch between openAI and Gemini API?

[Godrules500]

Can this work with the new streamUI from AI sdk 3.1?

[athrael-soju]

@Godrules500 I'd advice holding up, because the new AI SDK came with a few issues.

[Godrules500]

@Godrules500 I'd advice holding up, because the new AI SDK came with a few issues.

Bummer! Do you know what kind of issues off the top of your head? Also, wish there was a VertexAI implementation since I'm technically not allowed to use the google ai wrapper

[athrael-soju]

@Godrules500 I'd advice holding up, because the new AI SDK came with a few issues.

Bummer! Do you know what kind of issues off the top of your head? Also, wish there was a VertexAI implementation since I'm technically not allowed to use the google ai wrapper

This should give you an idea :) https://github.com/vercel/ai-chatbot/issues/325 Looks like it's a @vercel/ai related issue affecting several templates.