can_accept_as.RdDeclares the tool parameters and their classes. The agent is expected to only provide the arguments specified, and wizrd will attempt to convert whatever the agent sends to the declared class.
can_accept_as(`_x`, ..., `_parameters` = list(...))The Tool object
Each argument is any S7 class object, including base classes, S3 classes,
and unions. The arguments are
matched to the underlying function via match.call,
i.e. normal R calling semantics.
Convenience for providing a list instead of parameter classes
instead of individual args in ....
By default (if this function is never called), the formals of the
underlying function are taken as the parameters, and each accepts any
value. ..., if present, is declared as a list.
The agent is expected to send only the declared parameters, with some
potentially being captured by ... in the underlying
function. To request that the agent send extra arguments (as a list)
that would be captured by ..., pass an argument named
"\dots", e.g. accept_as(x, "..." = class_list), which is
the default behavior as mentioned above.
The arguments are evaluated in the env passed to
chat, so declaring an argument as a language object
(class_name or class_call) will evaluate the argument as
R code. See the below example for how this could be useful.
The input Tool object except with newly declared parameters and types.
if (FALSE) { # \dontrun{
var <- 1:10
agent <- llama() |>
equip(tool(mean) |> describe_with_Rd() |> can_accept_as(class_name)) |>
predict("What is the mean of var?")
} # }