Assistants - chat bots with skills

[cgillum]

This PR adds support for "assistants" - i.e., chat bots that can invoke functions as custom skills.

What's new

To support assistants, a new AssistantSkillTrigger is introduced. Any such functions in your project become automatically callable by chat bots. A sample demonstrating this is also included in the PR.

How it works

As part of this change, whenever a prompt is sent to a chat bot, the list of AssistantSkillTrigger functions are sent to the GPT language model. The model can then decide to respond by invoking one of those functions, which is done by the updated entity logic. The response is then sent back to the model so that it can respond to the user prompt appropriately.

(Image is borrowed from a Japanese presentation, introducing this idea)

Remaining work

• [x] An actual implementation: The original prototype depends on the Durable Functions implementation of chat bots. However, that's being replaced as part of this PR. Once merged, we can add the assistant functionality to the updated chat bot implementation.
• [ ] Support for more language stacks: Currently this feature only works with .NET in-proc. More work is needed to get it to work with .NET Isolated, Python, TypeScript, etc.
• [ ] Integration tests: We have an integration test for chat bots, but there should also be one for assistants with skills.

Future improvements

• Support for complex parameter types: Currently the skill triggers only support primitive types (string, int, float, bool, etc.).
• Fine-grained access controls for skills: Right now, all bots can call any skill defined in the function app. It might be worth considering whether certain skills can be restricted to specific assistants.
• User confirmation: Support marking certain skills as "requiring user confirmation" and then having the bot request permission before invoking that skill.

[cgillum]

A few other minor changes I made as part of this PR:

• Fixed several warnings in the existing code (nullability, etc.)
• Renamed some types to improve clarity
• Fixed inconsistencies in code style
• Refactored and simplified ChatBotService logic
• Added missing cancellation token flow to chat bot service logic
• Minor (breaking) schema changes to reduce data storage

[manvkaur]

We can add the pending tasks or future improvements to our internal board backlog

[cgillum]

BTW, here's what the conversation history looks like in Azure Table Storage after running the demo script.