Document strategies to handle invalid tool input

[verdie-g]

Type of issue

Missing information

Description

A LLM can provide invalid input when request tool calls. In this discussion with Stephen, several strategies were mentioned to deal with that.

The documentation could be improved with:

• FunctionInvokingChatClient.IncludeDetailedErrors to send serialization (and other) exceptions back to the LLM
• FunctionInvokingChatClient.FunctionInvoker to, for example, catch serialization exceptions. It could also be interesting to document that it can be used to implement a retry mechanism, like throwing a MyRetryException from a tool (because of incoherent input or something), catching it in the FunctionInvoker and returning a message to make the LLM retry "invalid something, please retry". Sharing good prompts that make some models retry a tool call would be a plus.
• strictJsonSchema special property for OpenAI

Page URL

https://learn.microsoft.com/en-us/dotnet/ai/quickstarts/use-function-calling

Content source URL

https://github.com/dotnet/docs/blob/main/docs/ai/quickstarts/use-function-calling.md

Document Version Independent Id

7228f348-6649-a87d-bae5-4a70817c03bb

Platform Id

5a517f05-b64b-c471-b99d-72253c213be1

Article author

@gewarren

Metadata

• ID: 27e2bbca-36b0-3883-11f7-321c044ee29a
• PlatformId: 5a517f05-b64b-c471-b99d-72253c213be1
• Service: dotnet
• Sub-service: intelligent-apps

[adegeo]

Hello. If there's a specific (or close to it) documentation you think needs updating, please navigate to the bottom of that documentation page and click the open issue button. Otherwise, we're not sure what article you think needs updating.

[verdie-g]

I updated the issue using the feedback template.