Expose internal ReqLLM data in action response

[sezaru]

Code of Conduct

• I agree to follow this project's Code of Conduct

AI Policy

• I agree to follow this project's AI Policy, or I agree that AI was not used while creating this issue.

Is your feature request related to a problem? Please describe.

Some times there are information that can only be retrieved from the ReqLLM data directly and that is discarded by AshAI.

For example, if I set a limit of tokens and want to check if the LLM generation was stopped by checking ReqLLM metadata field finish_reason.

Describe the solution you'd like

I'm not sure where this data should be stored, maybe in the context? But regardless on where, giving the user a way to retrieve it would be great, either by somehow returning it with the action return value, or, if that is not possible, having some way to handle that inside the action itself (in this case, it would be great if we could update the input data and request a retry (so, for example, in the case I mentioned above regarding max tokens, I would be able to check finish_reason, and, if the generating was stopped in the middle because of not enough tokens, I would be able to increase the amount of tokens and retry).

Describe alternatives you've considered

No response

Additional context

No response

[zachdaniel]

Yeah, we can do this via something like:

action :whatever, AshAi.Result do
  constraints [type: :integer, constraints: ...]
end

and if you do that, it will return %AshAi.ToolResult{result: ..., additional: ..., metadata: ...}

Then, when we call tools, if they return an %AshAi.Result{} we can unwrap it automatically?

---

Alternatively, we can add a feature of generic actions in Ash core like return_input?: true which causes it to return {:ok, result, %Ash.ActionInput{}}, and then ReqLLM can set that context on the action input some how?

[sezaru]

Do you have a preference on how to achieve this? If the change is simple, I can try doing it myself and creating a PR for it

[zachdaniel]

Hmm....I think I'd prefer the approach in Ash core to make it return the original input, and make it possible to return an updated input from the generic action run function.