Utilizing multiple function/tools calls?

[devjn]

Can we use function calls instead of JSON mode? Particularly would be nice to get example with Vertex AI Gemini as they don't support Union types.

Currently I can see that when specified Mode.VERTEXAI_TOOLS for instructor, and then for BaseAgentConfig(output_schema=AgentOutputSchema)
I can see from logs that 1 tool is defined like this:
'tools': [function_declarations {
name: "AgentOutputSchema
...
}]

How could we define more tools and instruct correctly?

[KennyVaneetvelde]

Hello, unless I misunderstand your question, I believe this question is more suited for https://github.com/instructor-ai/instructor

Maybe though it's not possible with Vertex AI in tool mode, either way I'm sure the folks at Instructor are more knowledgeable on this specific subject!

[devjn]

Yeah gemini question seems to be more for instructor. However, currently there is parallel-function-calling that takes Iterable so to utilize it we would need output_schema to allow it. Something like this: BaseAgentConfig(output_schema=Iterable[WeatherSchema | GoogleSearchSchema])

[KennyVaneetvelde]

@devjn Have you tried this? Anything that works for instructor should work within atomic agents, or is there something specific that's not working or holding you back here?

[devjn]

@KennyVaneetvelde The issue currently is that BaseAgent expects BaseAgentConfig for output_schema and if I use Iterable I get: ValidationError: 1 validation error for BaseAgentConfig output_schema Input should be a subclass of BaseModel
Or should I not use BaseAgent?

[KennyVaneetvelde]

Right, The config is just for configuration, but the point still stands as right now it only supports a class that extends from Pydantic's BaseModel but I get what you mean, you need it to accept an iterable...

I will try to look into that!