Rand Stats

The uploading author of cpan:BDUGGAN does not match the META author of zef:bduggan.

LLM::DWIM

cpan:BDUGGAN

Actions Status Actions Status

NAME

LLM::DWIM -- Do What I Mean, with help from large language models.

SYNOPSIS

use LLM::DWIM;

say dwim "How many miles is it from the earth to the moon?";
# Approximately 238,900 miles (384,400 kilometers)

say dwim "@NothingElse How many miles is it from the earth to the moon? #NumericOnly";
# 238900

sub distance-between($from,$to) {
  dwim "@NothingElse #NumericOnly What is the distance in miles between $from and $to?";
}

say distance-between("earth","sun");
# 92955887.6 miles

Meanwhile, in ~/.config/llm-dwim.toml:

evaluator = "gemini"
gemini.temperature = 0.5

DESCRIPTION

This is a simple wrapper around LLM::Functions, and LLM::Prompts It provides a single subroutine, dwim, that sends a string to an LLM evaluator, making use of a configuration file to say a little more about what you mean.

FUNCTIONS

dwim

sub dwim(Str $str) returns Mu

This function takes a string, expands it using LLM::Prompts, and uses LLM::Functions to evaluate the string.

It is mostly equivalent to:

use LLM::Functions;
use LLM::Prompts;
use TOML;

my $conf-dir = %*ENV<XDG_HOME> // $*HOME.child('.config');
my $conf = from-toml($conf-dir.child('llm-dwim.toml').IO.slurp);
my $evaluator = $conf<evaluator>;
my &evaluator //= llm-function(
    llm-evaluator => llm-configuration( $evaluator, |%( $conf{ $evaluator } ) )
);
my $msg = llm-prompt-expand($str);
evaluator($msg);

For diagnostics, use Log::Async and add a tap, like so:

use LLM::DWIM;
use Log::Async;

logger.send-to($*ERR);

say dwim "How many miles is it from earth is the moon? #NumericOnly";

CONFIGURATION

This module looks for llm-dwim.toml in either XDG_HOME or HOME/.config. This can be overridden by setting DWIM_LLM_CONF to another filename.

The configuration file should be in TOML format and should contain at least one key, evaluator, which should be the name of the LLM evaluator to use. Evaluators can be configured using TOML syntax, with the evaluator name as the key.

Sample configurations:

Use Gemini (which has a free tier) :

evaluator = "gemini"

Use OpenAI, and modify some parameters:

evaluator = "OpenAI"
OpenAI.temperature = 0.9
OpenAI.max-tokens = 100

See LLM::Functions for all of the configuration options.

COMMAND LINE USAGE

Also, this package includes a llm-dwim script that can be used to evaluate a string from the command line. It will look for the configuration file in the same way as the module. It accepts either a string or a list of words. The single argument "-" will cause it to read from standard input.

Sample usage:

llm-dwim -h  # get usage
llm-dwim "How many miles is it from the earth to the moon?"
llm-dwim -v how far is it from the earth to the moon\?
echo "what is the airspeed velocity of an unladen swallow?" | llm-dwim -

SEE ALSO

LLM::Functions, LLM::Prompts

This was inspired by the also excellent DWIM::Block module.

AUTHOR

Brian Duggan