Create a provider-level prompt wrap

Build a provider-specific prompt wrap, to store on an llm_provider object (with $add_prompt_wrap()). These prompt wraps can be applied before or after any prompt-specific prompt wraps. In this way, you can ensure that certain prompt wraps are always applied when using a specific LLM provider.

Usage

provider_prompt_wrap(
  modify_fn = NULL,
  extraction_fn = NULL,
  validation_fn = NULL,
  handler_fn = NULL,
  parameter_fn = NULL,
  type = c("unspecified", "mode", "tool", "break", "check"),
  name = NULL
)

Arguments

modify_fn

A function that takes the previous prompt text (as first argument) and returns the new prompt text

extraction_fn

A function that takes the LLM response (as first argument) and attempts to extract a value from it. Upon succesful extraction, the function should return the extracted value. If the extraction fails, the function should return a llm_feedback() message to initiate a retry. A llm_break() can be returned to break the extraction and validation loop, ending send_prompt()

validation_fn

A function that takes the (extracted) LLM response (as first argument) and attempts to validate it. Upon succesful validation, the function should return TRUE. If the validation fails, the function should return a llm_feedback() message to initiate a retry. A llm_break() can be returned to break the extraction and validation loop, ending send_prompt()

handler_fn

A function that takes a 'completion' object (a result of a request to a LLM, as returned by $complete_chat() of a llm_provider object) as first argument and the llm_provider object as second argument. The function should return a (modified or identical) completion object. This can be used for advanced side effects, like logging, or native tool calling, or keeping track of token usage. See llm_provider for more information; handler_fn is attached to the llm_provider object that is being used. When using an llm_provider_ellmer(), the up-to-date ellmer_chat is synced onto the provider before handlers run. This allows handlers to access, for instance, the current cost of the conversation, and, for instance, to stop the conversation if a certain budget is exceeded. For example usage, see source code of answer_using_tools()

parameter_fn

A function that takes the llm_provider object which is being used with send_prompt() and returns a named list of parameters to be set in the llm_provider object via its $set_parameters() method. This can be used to configure specific parameters of the llm_provider object when evaluating the prompt. For example, answer_as_json() may set different parameters for different APIs related to JSON output. This function is typically only used with advanced prompt wraps that require specific settings in the llm_provider object

type

The type of prompt wrap. Must be one of:

• "unspecified": The default type, typically used for prompt wraps which request a specific format of the LLM response, like answer_as_integer()

• "mode": For prompt wraps that change how the LLM should answer the prompt, like answer_by_chain_of_thought() or answer_by_react()

• "tool": For prompt wraps that enable the LLM to use tools, like answer_using_tools() or answer_using_r() when 'output_as_tool' = TRUE

• "break": For prompt wraps that may break the extraction and validation loop, like quit_if(). These are applied before type "unspecified" as they may instruct the LLM to not answer the prompt in the manner specified by those prompt wraps

• "check": For prompt wraps that apply a last check to the final answer, after all other prompt wraps have been evaluated. These prompt wraps may only contain a validation function, and are applied after all other prompt wraps have been evaluated. These prompt wraps are even applied after an earlier prompt wrap has broken the extraction and validation loop with llm_break()

Types are used to determine the order in which prompt wraps are applied. When constructing the prompt text, prompt wraps are applied to the base prompt in the following order: 'check', 'unspecified', 'break', 'mode', 'tool'. When evaluating the LLM response and applying extraction and validation functions, prompt wraps are applied in the reverse order: 'tool', 'mode', 'break', 'unspecified', 'check'. Order among the same type is preserved in the order they were added to the prompt.

name

An optional name for the prompt wrap. This can be used to identify the prompt wrap in the tidyprompt object

Value

A provider_prompt_wrap object, to be stored on an llm_provider object

Examples

ollama <- llm_provider_ollama()

# Add a "short answer" mode (provider-level post prompt wrap)
ollama$add_prompt_wrap(
  provider_prompt_wrap(
    modify_fn = \(txt) paste0(
      txt,
      "\n\nPlease answer concisely (< 2 sentences)."
    )
  ),
  position = "post"
)

# Use as usual: wraps are applied automatically
if (FALSE) { # \dontrun{
"What's a vignette in R?" |> send_prompt(ollama)
} # }