LLMFunction

✖
`LLMFunction`

[Experimental]

This functionality requires LLM access »

✖

LLMFunction[prompt]

represents a template for a large language model (LLM) prompt.

✖

LLMFunction[{prompt₁,prompt₂,…}]

represents a combination of multiple prompts.

✖

LLMFunction[prompt,form]

includes the interpreter form to apply to the response.

✖

LLMFunction[…][params]

give the LLM service response for prompt applied to parameters params.

Details and Options

An LLMFunction can be used to generate text using a large language model (LLM). It can create content, complete sentences, extract information and more.
LLMFunction requires external service authentication, billing and internet connectivity.
The prompt_i supports the following values:

	"text"	static text
	LLMPrompt["name"]	a repository prompt
	StringTemplate[…]	templated text
	TemplateObject[…]	template for creating a prompt
	{prompt₁,…}	a list of prompts

Prompts created with TemplateObject can contain text and images. Not every LLM supports image input.
LLMFunction supports the following options:

InsertionFunction	TextString	function or format to apply before inserting expressions
CombinerFunction	StringJoin	function to apply to combine pieces within a prompt
Authentication	Automatic	explicit user ID and API key
LLMEvaluator	$LLMEvaluator	LLM configuration to use

LLMEvaluator can be set to an LLMConfiguration object or an association with any of the following keys:

	"MaxTokens"	maximum amount of tokens to generate
	"Model"	base model
	"PromptDelimiter"	string to insert between prompts
	"Prompts"	initial prompts
	"StopTokens"	tokens on which to stop generation
	"Temperature"	sampling temperature
	"ToolMethod"	method to use for tool calling
	"Tools"	list of LLMTool objects to make available
	"TopProbabilities"	sampling classes cutoff
	"TotalProbabilityCutoff"	sampling probability cutoff (nucleus sampling)

Valid forms of "Model" include:
name named model

{service,name} named model from service

<|"Service"service,"Name"name,"Task"task|> fully specified model
The generated text is sampled from a distribution. Details of the sampling can be specified using the following properties of LLMEvaluator:

"Temperature"t	Automatic	sample using a positive temperature t
"TopProbabilities"k	Automatic	sample only among the k highest-probability classes
"TotalProbabilityCutoff"p	Automatic	sample among the most probable choices with an accumulated probability of at least p (nucleus sampling)

The setting "Temperature"Automatic resolves to zero temperature within LLMFunction. The other parameters use the default for the specified "Model".
Multiple prompts are separated by the "PromptDelimiter" property of the LLMEvaluator.
Possible values for Authentication are:

	Automatic	choose the authentication scheme automatically
	Environment	check for a key in the environment variables
	SystemCredential	check for a key in the system keychain
	ServiceObject[…]	inherit the authentication from a service object
	assoc	provide an explicit key and user ID

With AuthenticationAutomatic, the function checks the variable ToUpperCase[service]<>"_API_KEY" in Environment and SystemCredential; otherwise, it uses ServiceConnect[service].
When using Authenticationassoc, assoc can contain the following keys:
"ID" user identity

"APIKey" API key used to authenticate
LLMFunction uses machine learning. Its methods, training sets and biases included therein may change and yield varied results in different versions of the Wolfram Language.