--- title: "LLMPromptGenerator" language: "en" type: "Symbol" summary: "LLMPromptGenerator[f] represents a prompt generator that uses the function f. LLMPromptGenerator[f, inputspec] provides the specified inputspec to f." keywords: - prompt generation - prompt generator - dynamic prompt generation - system prompt generation - message-based prompt - dynamic prompt - contextual prompt - LLM - LLM prompt function - LLM input function - LLM function - prompt customization - prompt templating - conversation context - chat history - semantic prompt - semantic search - image prompt - multimodal prompt - text-to-image - prompt output - repository prompt - static prompt - generated prompt - prompt specification - input mapping - association-based input - LLM customization - Wolfram LLM - LLM prompt pipeline - function-based prompt - AI assistant input - template prompt creation - chat model input - prompt programming - structured LLM input - LLM context function - contextual generation - LLM prompt logic - Wolfram AI prompt - prompt result - multi-input prompt - chat-enhanced prompt - prompt framework - retrieval-augmented generation - RAG - semantic retrieval - retrieval prompt - search-enhanced prompt - document-based prompt - external knowledge injection - knowledge grounding - retrieval pipeline - RAG implementation - LLM with search canonical_url: "https://reference.wolfram.com/language/ref/LLMPromptGenerator.html" source: "Wolfram Language Documentation" related_guides: - title: "LLM-Related Functionality" link: "https://reference.wolfram.com/language/guide/LLMFunctions.en.md" - title: "Natural Language Processing" link: "https://reference.wolfram.com/language/guide/NaturalLanguageProcessing.en.md" related_functions: - title: "LLMSynthesize" link: "https://reference.wolfram.com/language/ref/LLMSynthesize.en.md" - title: "LLMFunction" link: "https://reference.wolfram.com/language/ref/LLMFunction.en.md" - title: "ChatEvaluate" link: "https://reference.wolfram.com/language/ref/ChatEvaluate.en.md" - title: "SemanticSearch" link: "https://reference.wolfram.com/language/ref/SemanticSearch.en.md" - title: "SemanticSearchIndex" link: "https://reference.wolfram.com/language/ref/SemanticSearchIndex.en.md" - title: "LLMPrompt" link: "https://reference.wolfram.com/language/ref/LLMPrompt.en.md" - title: "LLMTool" link: "https://reference.wolfram.com/language/ref/LLMTool.en.md" --- [EXPERIMENTAL] # LLMPromptGenerator LLMPromptGenerator[f] represents a prompt generator that uses the function f. LLMPromptGenerator[f, inputspec] provides the specified inputspec to f. ## Details * ``LLMPromptGenerator`` is used to add message-dependent context to an LLM prompt. * ``LLMPromptGenerator`` can be used in retrieval-augmented generation (RAG) workflows to construct prompts dynamically using retrieved documents, semantic indexes or other external knowledge sources. * Possible values for ``inputspec`` are: | | | | -------------- | --------------------------------------- | | "Input" | the last user input (default) | | "Messages" | the list of messages | | "LLMEvaluator" | the current LLMConfiguration[…] | | "ChatObject" | the whole conversation as ChatObject[…] | | {spec1, …} | an association of inputs | * The result of ``f`` has to be a valid prompt. Possible values include: | | | | ---------------------- | ------------------------------ | | "text" | static text | | LLMPrompt["name"] | a repository prompt | | StringTemplate[…] | templated text | | TemplateObject[…] | template for creating a prompt | | Image[…] | an image | | SemanticSearchIndex[…] | a semantic search index | | {prompt1, …} | a list of prompts | * Template objects are automatically converted to strings via ``TemplateObject[…][]``. * A prompt created with ``TemplateObject`` can contain text and images. * Not every LLM supports image input. * If the result of ``f`` is not a string, a list of strings or an image, it will be converted using ``TextString``. ## Examples (7) ### Basic Examples (1) Define a prompt generator that always prepends the current evaluation date and time: ```wl In[1]:= gen = LLMPromptGenerator[{"Current date/time: ", DateString[]}&] Out[1]= LLMPromptGenerator[Association["InputSpecifications" -> {"Input"}, "Function" -> ({"Current date/time: ", DateString[]} & ), "InputSelector" -> Lookup["Input"], "Options" -> {}, "LLMPacletVersion" -> "1.2.10"]] ``` Test the generator: ```wl In[2]:= gen["hello there"] Out[2]= {"Current date/time: ", "Mon 24 Jun 2024 18:31:48"} ``` Use the generator in an interaction with an LLM: ```wl In[3]:= LLMSynthesize["what day is today?", LLMEvaluator -> <|"Prompts" -> gen|>] Out[3]= "Today is Monday, June 24, 2024." ``` ### Scope (5) #### Function (3) Use a simple prompt: ```wl In[1]:= LLMSynthesize["hi", LLMEvaluator -> <|"Prompts" -> LLMPromptGenerator@LLMPrompt["R2D2"]|>] Out[1]= "Beep-beep!" ``` --- Use a function to define the tone of the answer at evaluation time based on the input: ```wl In[1]:= gen = LLMPromptGenerator["Talk like a " <> First[StringCases[#, StartOfString ~~ "@" ~~ role : (LetterCharacter..) ~~ WordBoundary :> role], "pirate"]&] Out[1]= LLMPromptGenerator[Association["InputSpecifications" -> {"Input"}, "Function" -> (StringJoin["Talk like a ", First[StringCases[#1, StartOfString~~"@"~~role:LetterCharacter..~~WordBoundary :> role], "pirate"]] & ), "InputSelector" -> Lookup["Input"], "Options" -> {}, "LLMPacletVersion" -> "1.2.10"]] ``` Try with a cowboy role: ```wl In[2]:= LLMSynthesize["@cowboy hi!", LLMEvaluator -> <|"Prompts" -> gen|>] Out[2]= "Well, howdy there, partner! What brings ya to this neck of the digital woods?" ``` --- Define a generator on a ``SemanticSearchIndex`` : ```wl In[1]:= gen = LLMPromptGenerator@CreateSemanticSearchIndex[WikipediaData["light"]] Out[1]= LLMPromptGenerator[Association["InputSpecifications" -> {"Input"}, "Function" -> TemplateObject[{"Answer using this information:", "\n\n", "BEGIN CONTEXT", "\n\n", TemplateSlot["Items", InsertionFunction -> (StringRiffle[#1, Wolf ... "GeneratedAssetLocation" -> "LocalObject", "TagKeys" -> {}, "Version" -> 1, "Hash" -> 103221403795781887489925846881430479734]], #1, "Items"]] & ), "InputSelector" -> Lookup["Input"], "Options" -> {}, "LLMPacletVersion" -> "1.2.10"]] ``` Use it to add content semantically relevant to the query: ```wl In[2]:= LLMSynthesize["What is the connection between light and Aphrodite in one sentence", LLMEvaluator -> <|"Prompts" -> gen|>] Out[2]= "According to Empedocles, the goddess Aphrodite created the human eye from the four elements and lit a fire within it, enabling sight through the interaction of this light with external rays." ``` Compare with the uninformed LLM answer: ```wl In[3]:= LLMSynthesize["What is the connection between light and Aphrodite in one sentence"] Out[3]= "The connection between light and Aphrodite lies in her association with beauty and radiance, often symbolized by the light of the morning star (Venus), which is a celestial embodiment of the goddess." ``` #### Input Specifications (2) Define a generator that echoes the calling function input (default argument): ```wl In[1]:= LLMSynthesize["this is the input", "PromptText", LLMEvaluator -> <|"Prompts" -> LLMPromptGenerator[Function[Echo[#]; "the prompt"]]|>] >> "this is the input" Out[1]= "the prompt this is the input" ``` --- Give a custom input specification: ```wl In[1]:= LLMSynthesize["This is the input", "PromptText", LLMEvaluator -> <|"Prompts" -> LLMPromptGenerator[Function[StringRiffle[{"* User input is: \"" <> ToString[#Input] <> "\"", "* Current LLM model is: " <> #LLMEvaluator["Model"]["Name"]}, " "]], {"LLMEvaluator", "Input"}]|>] Out[1]= "* User input is: \"This is the input\" * Current LLM model is: gpt-4o This is the input" ``` ### Possible Issues (1) A list of prompts does not evaluate when applied to the input: ```wl In[1]:= LLMPromptGenerator[{"Current date/time: ", DateString[]}]["content"] Out[1]= Failure[LLMPromptGenerator, Association["MessageTemplate" :> LLMPromptGenerator::uneval, "MessageParameters" :> {InputForm[{"Current date/time: ", "Thu 27 Jun 2024 15:36:27"}], InputForm["content"]}]] ``` Wrap the list using ``TemplateObject`` or ``Function`` : ```wl In[2]:= LLMSynthesize["content", "PromptText", LLMEvaluator -> {"Prompts" -> LLMPromptGenerator[TemplateObject[{"Current date/time: ", TemplateExpression@DateString[]}]]}] Out[2]= "Current date/time: Thu 27 Jun 2024 15:36:35 content" ``` ## See Also * [`LLMSynthesize`](https://reference.wolfram.com/language/ref/LLMSynthesize.en.md) * [`LLMFunction`](https://reference.wolfram.com/language/ref/LLMFunction.en.md) * [`ChatEvaluate`](https://reference.wolfram.com/language/ref/ChatEvaluate.en.md) * [`SemanticSearch`](https://reference.wolfram.com/language/ref/SemanticSearch.en.md) * [`SemanticSearchIndex`](https://reference.wolfram.com/language/ref/SemanticSearchIndex.en.md) * [`LLMPrompt`](https://reference.wolfram.com/language/ref/LLMPrompt.en.md) * [`LLMTool`](https://reference.wolfram.com/language/ref/LLMTool.en.md) ## Related Guides * [LLM-Related Functionality](https://reference.wolfram.com/language/guide/LLMFunctions.en.md) * [Natural Language Processing](https://reference.wolfram.com/language/guide/NaturalLanguageProcessing.en.md) ## History * [Introduced in 2024 (14.1)](https://reference.wolfram.com/language/guide/SummaryOfNewFeaturesIn141.en.md)