Phi-Silica: JSON Mode, Function calling

[purudpd]

Over the past weekend, I played around with Phi-Silica on experimental WindowsApp SDK. It's very promising and impresses with it's efficiency and speed, running on the local NPU.

This would be great for small NLP tasks (sentiment, decision making etc.) but I also have some feedback:

• No JSON mode: Difficult the adhere the model to a particular JSON format
• Support for function calling: Is there any potential here?
• Knowledge Cut-off: The internal knowledge seems to be quite outdated.
• Potentially bigger context window?

Finally, are there any plans to have more models available (model catalog) or allow developers to bring their own models?

[connorMicrosoft]

Thank you! We are tracking this as a feature request on our end.

[purudpd]

With the release of Phi-4 mini and Phi-4 instruct multi-modal, can we see these being available through the Windows App SDK in the near future?

[connorMicrosoft]

Can you clarify the challenges you're running into with JSON? What scenario are you trying to enable?

We're not able to share updates on our future roadmap right now, but we'll keep you posted!

[purudpd]

@connorMicrosoft Challenges with JSON:
The model does quite poorly with instruction following and outputting valid JSONs.

• It tends to add additional text, before or after the JSON output.

I was trying to enable tool or function calling on a winui3 app, allowing the SLM to carry out tasks. This involved:

• Deciding from a few "tools" -> Outputting the decision as a valid JSON -> Parsing the JSON to perform the operation.

Unfortunately, the model does quite poorly for any such scenarios.