prompt_as.RdConfigure an agent to be prompted in a systematic way.
prompt_as(x, format, ...)
system_prompt_as(x, format, ...)The Agent object
A TextFormat object, or something convertible to one, representing the prompt format. A function is converted to a TextFormat that expects a list of arguments to be passed as input. The arguments are passed to the function, and the return value is sent to the model. If this argument is a string, the system will try to interpret it as either a file, glue format string or Langsmith Hub ID. If it resembles a file, the file is read as a string and treated as a glue template. If it is a Langsmith Hub ID, the corresponding prompt template is downloaded and set on the agent (including the system prompt). Both fstring (via glue) and mustache (via whisker) templates are supported from the hub.
Arguments passed to methods
The specified format object controls (via dispatch) how the
user input is converted to a textual prompt. In most cases, templating
is sufficient, and the user can pass template parameters as a list (or
similar object, like an environment) as the input argument to
e.g. chat. An example of a more complex conversion is
Retrieval Augmented Generation (RAG), via rag_with.
When passing Langsmith Hub IDs, note that prompt_as is
considered the more general entry point, supporting chat and
structured output prompt templates in addition to basic prompt
templates. system_prompt_as only supports basic prompt
templates, since it can only modify the system prompt.
The input Agent object except with a new prompt format.
if (FALSE) { # \dontrun{
llama() |>
prompt_as("Output a {nchar}-letter word starting with '{initial}'.") |>
predict(list(nchar = 4L, initial = "a"))
llama() |>
system_prompt_as("Answer questions about {language}") |>
chat(system_params = list(language = "R")) |>
predict("Who created the language?")
} # }