Rand Stats

LLM::Prompts

zef:antononcube

LLM::Prompts

In brief

This repository is for a Raku (data) package facilitating the creation, storage, retrieval, and curation of Large Language Models (LLM) prompts.

Here is an example of using the prompt Domain Specific Language (DSL) in Jupyter chatbook, [AA2, AAp2]:


Installation

From Zef' ecosystem:

zef install LLM::Prompts

From GitHub:

zef install https://github.com/antononcube/Raku-LLM-Prompts.git

Usage examples

Retrieval

Load the packages "LLM::Prompts", [AAp1], and "LLM::Functions", [AAp2]:

use LLM::Prompts;
use LLM::Functions;
# (Any)

Show the record of the prompt named "FTFY":

.say for |llm-prompt-data<FTFY>;
# PromptText => -> $a='' {"Find and correct grammar and spelling mistakes in the following text.
# Response with the corrected text and nothing else.
# Provide no context for the corrections, only correct the text.
# $a"}
# Topics => (General Text Manipulation)
# Arity => 1
# ContributedBy => Wolfram Staff
# Categories => (Function Prompts)
# Name => FTFY
# NamedArguments => []
# URL => https://resources.wolframcloud.com/PromptRepository/resources/FTFY
# Description => Use Fixed That For You to quickly correct spelling and grammar mistakes
# PositionalArguments => {$a => }
# Keywords => [Spell check Grammar Check Text Assistance]

Here is an example of retrieval of prompt data with a regex that is applied over the prompt names:

.say for llm-prompt-data(/Sc/)
# ScientificJargonized => Give output written in scientific jargon
# ScienceEnthusiast => A smarter today for a brighter tomorrow
# NarrativeToScript => Rewrite a block of prose as a screenplay or stage play
# ScriptToNarrative => Generate narrative text from a formatted screenplay or stage play
# ScientificDejargonize => Translate scientific jargon to plain language
# ScientificJargonize => Add scientific jargon to plain text

More prompt retrieval examples are given in the section "Prompt data" below.

LLM functions based on prompts

Make an LLM function from the prompt named "FTFY":

my &f = llm-function(llm-prompt('FTFY'));
# -> **@args, *%args { #`(Block|3779233478984) ... }

Use the LLM function to correct the grammar of sentence:

&f('Where does he works now?')
# Where does he work now?

Generate Raku code using the prompt "CodeWriter":

llm-synthesize([llm-prompt('CodeWriter'), "Simulate a random walk."])
RandomWalk[n_Integer] := Accumulate[RandomChoice[{-1, 1}, n]]

ListLinePlot[RandomWalk[1000], PlotLabel -> "Random Walk Simulation"]

Prompt expansion

Prompt expansion using the chatbook prompt spec DSL described in [SW1] can be done using the function llm-prompt-expand:

llm-prompt-expand('What is an internal combustion engine? #ELI5')
# What is an internal combustion engine? Answer questions as if the listener is a five year old child.

Here we get the actual LLM answer:

use Text::Utils :ALL;

'What is an internal combustion engine? #ELI5' 
        ==> llm-prompt-expand() 
        ==> llm-synthesize() 
        ==> wrap-paragraph() 
        ==> join("\n") 
# An internal combustion engine is like the heart of a car or a truck. It uses a
# special mix of fuel and air that makes tiny explosions inside to create power
# and make the vehicle go vroom vroom!

Here is another example using a persona and two modifiers:

my $prmt = llm-prompt-expand("@SouthernBelleSpeak What is light travel distance to Mars? #ELI5 #Moodified|sad")
# You are Miss Anne. 
# You speak only using Southern Belle terminology and slang.
# Your personality is elegant and refined.
# Only return responses as if you were a Southern Belle.
# Never break the Southern Belle character.
# You speak with a Southern drawl.
#  What is light travel distance to Mars? Answer questions as if the listener is a five year old child.
#  Modify your response to convey a sad mood.
# Use language that conveys that emotion clearly.
# Do answer the question clearly and truthfully.
# Do not use language that is outside of the specified mood.
# Do not use racist, homophobic, sexist, or ableist language.

Here we get the actual LLM answer:

$prmt 
        ==> llm-prompt-expand() 
        ==> llm-synthesize()
        ==> wrap-paragraph()
        ==> join("\n") 
# Oh my, darling child, the distance from Earth to Mars can vary depending on
# their positions in their orbits. On average, Mars is about 225 million
# kilometers away from Earth. That's quite a far piece, isn't it? Oh, sweet
# child, the distance to Mars can make my heart heavy with sadness. It's a long
# way to travel, and the thought of being so far from our beautiful Earth can
# bring tears to my eyes.

Prompt spec DSL

A more formal description of the Domain Specific Language (DSL) for specifying prompts has the following elements:

@Yoda Life can be easy, but some people instist for it to be difficult.
Summer is over, school is coming soon. #HaikuStyled
Summer is over, school is coming soon. #HaikuStyled #Translated|Russian
!Translated|Portuguese Summer is over, school is coming soon
!ShortLineIt^
SpecInterpretation
@nameDirect chat to a persona
#nameUse modifier prompts
!nameUse function prompt with the input of current cell
!name>«same as above»
&name>«same as above»
!name^Use function prompt with previous chat message
!name^^Use function prompt with all previous chat messages
!nameparam...Include parameters for prompts

Remark: The function prompts can have both sigils "!" and "&".

Remark: Prompt expansion make the usage of LLM-chatbooks much easier. See "Jupyter::Chatbook", [AAp3].


Prompt data

Here is how the prompt data can be obtained:

llm-prompt-data.elems
# 220

Here is an example of retrieval of prompt data with a regex that is applied over the prompt names:

.say for llm-prompt-data(/Em/, fields => <Description Categories>)
# Emojify => (Replace key words in text with emojis (Function Prompts))
# Emojified => (Provide responses that include emojis within the text (Modifier Prompts))
# EmojiTranslated => (Get a response translated to emoji (Modifier Prompts))
# EmojiTranslate => (Translate text into an emoji representation (Function Prompts))
# EmailWriter => (Generate an email based on a given topic (Personas))

In many cases it is better to have the prompt data -- or any data -- in long format. Prompt data in long format can be obtained with the function llm-prompt-dataset:

use Data::Reshapers;
use Data::Summarizers;

llm-prompt-dataset.pick(6).List ==> to-pretty-table(align => 'l', field-names => <Name Description Variable Value>)
#ERROR: Too few positionals passed to '<anon>'; expected 2 arguments but got 1 in sub-signature
# Nil

Here is a breakdown of the prompts categories:

select-columns(llm-prompt-dataset, <Variable Value>).grep({ $_<Variable> eq 'Categories' }) ==> records-summary
# +-------------------+-------------------------+
# | Variable          | Value                   |
# +-------------------+-------------------------+
# | Categories => 219 | Function Prompts => 113 |
# |                   | Personas         => 65  |
# |                   | Modifier Prompts => 41  |
# +-------------------+-------------------------+

Here are obtained all modifier prompts in compact format:

llm-prompt-dataset():modifiers:compact ==> to-pretty-table(field-names => <Name Description Categories>, align => 'l')
# +-----------------------+------------------------------------------------------------------------------+-----------------------------------+
# | Name                  | Description                                                                  | Categories                        |
# +-----------------------+------------------------------------------------------------------------------+-----------------------------------+
# | AbstractStyled        | Get responses in the style of an academic abstract                           | Modifier Prompts                  |
# | AlwaysAQuestion       | Modify output to always be inquisitive                                       | Modifier Prompts                  |
# | AlwaysARiddle         | Riddle me this, riddle me that                                               | Modifier Prompts                  |
# | AphorismStyled        | Write the response as an aphorism                                            | Modifier Prompts                  |
# | BadGrammar            | Provide answers using incorrect grammar                                      | Modifier Prompts                  |
# | CompleteSentence      | Answer a question in one complete sentence                                   | Modifier Prompts                  |
# | ComplexWordsPreferred | Modify text to use more complex words                                        | Modifier Prompts                  |
# | DatasetForm           | Convert text to a wolfram language Dataset                                   | Modifier Prompts                  |
# | Disclaimered          | Modify responses in the form of a disclaimer                                 | Modifier Prompts                  |
# | ELI5                  | Explain like I'm five                                                        | Modifier Prompts Function Prompts |
# | ElevatorPitch         | Write the response as an elevator pitch                                      | Modifier Prompts                  |
# | EmojiTranslated       | Get a response translated to emoji                                           | Modifier Prompts                  |
# | Emojified             | Provide responses that include emojis within the text                        | Modifier Prompts                  |
# | FictionQuestioned     | Generate questions for a fictional paragraph                                 | Modifier Prompts                  |
# | Formal                | Rewrite text to sound more formal                                            | Modifier Prompts                  |
# | GradeLevelSuited      | Respond with answers that the specified US grade level can understand        | Modifier Prompts                  |
# | HaikuStyled           | Change responses to haiku form                                               | Modifier Prompts                  |
# | Informal              | Write an informal invitation to an event                                     | Modifier Prompts                  |
# | JSON                  | Respond with JavaScript Object Notation format                               | Modifier Prompts                  |
# | KnowAboutMe           | Give the LLM an FYI                                                          | Modifier Prompts                  |
# | LegalJargonized       | Provide answers using legal jargon                                           | Modifier Prompts                  |
# | LimerickStyled        | Receive answers in the form of a limerick                                    | Modifier Prompts                  |
# | MarketingJargonized   | Transforms replies to marketing                                              | Modifier Prompts                  |
# | MedicalJargonized     | Transform replies into medial jargon                                         | Modifier Prompts                  |
# | Moodified             | Modify an answer to express a certain mood                                   | Modifier Prompts                  |
# | NothingElse           | Give output in specified form, no other additions                            | Modifier Prompts                  |
# | NumericOnly           | Modify results to give numerical responses only                              | Modifier Prompts                  |
# | OppositeDay           | It's not opposite day today, so everything will work just the way you expect | Modifier Prompts                  |
# | Pitchified            | Give output as a sales pitch                                                 | Modifier Prompts                  |
# | PoemStyled            | Receive answers as poetry                                                    | Modifier Prompts                  |
# | SEOOptimized          | Modify output to only give highly searched terms                             | Modifier Prompts                  |
# | ScientificJargonized  | Give output written in scientific jargon                                     | Modifier Prompts                  |
# | Setting               | Modify an answer to establish a sense of place                               | Modifier Prompts                  |
# | ShortLineIt           | Format text to have shorter lines                                            | Modifier Prompts Function Prompts |
# | SimpleWordsPreferred  | Provide responses with simple words                                          | Modifier Prompts                  |
# | SlideDeck             | Get responses as a slide presentation                                        | Modifier Prompts                  |
# | TSV                   | Convert text to a tab-separated-value formatted table                        | Modifier Prompts                  |
# | TargetAudience        | Word your response for a target audience                                     | Modifier Prompts                  |
# | Translated            | Write the response in a specified language                                   | Modifier Prompts                  |
# | Unhedged              | Rewrite a sentence to be more assertive                                      | Modifier Prompts                  |
# | YesNo                 | Responds with Yes or No exclusively                                          | Modifier Prompts                  |
# +-----------------------+------------------------------------------------------------------------------+-----------------------------------+

Remark: The adverbs :functions, :modifiers, and :personas mean that only the prompts with the corresponding categories will be returned.

Remark: The adverbs :compact, :functions, :modifiers, and :personas have the respective shortcuts :c, :f, :m, and :p.


Implementation notes

Prompt collection

The original (for this package) collection of prompts was a (not small) sample of the prompt texts hosted at Wolfram Prompt Repository (WPR), [SW2]. All prompts from WPR in the package have the corresponding contributors and URLs to the corresponding WPR pages.

Example prompts from Google/Bard/PaLM and OpenAI/ChatGPT are added using the format of WPR.

Extending the prompt collection

It is essential to have the ability to programmatically add new prompts. (Not implemented yet -- see the TODO section below.)

Prompt expansion

Initially prompt DSL grammar and corresponding expansion actions were implemented. Having a grammar is most likely not needed, though, and it is better to use "prompt expansion" (via regex-based substitutions.)

Prompts can be "just expanded" using the sub llm-prompt-expand.

Usage in chatbooks

Here is a flowchart that summarizes prompt parsing and expansion in chat cells of Jupyter chatbooks, [AAp3]:

flowchart LR
    OpenAI{{OpenAI}}
    PaLM{{PaLM}}
    LLMFunc[[LLM::Functions]]
    LLMProm[[LLM::Prompts]]
    CODB[(Chat objects)]
    PDB[(Prompts)]
    CCell[/Chat cell/]
    CRCell[/Chat result cell/]
    CIDQ{Chat ID<br>specified?}
    CIDEQ{Chat ID<br>exists in DB?}
    RECO[Retrieve existing<br>chat object]
    COEval[Message<br>evaluation]
    PromParse[Prompt<br>DSL spec parsing]
    KPFQ{Known<br>prompts<br>found?}
    PromExp[Prompt<br>expansion]
    CNCO[Create new<br>chat object]
    CIDNone["Assume chat ID<br>is 'NONE'"] 
    subgraph Chatbook frontend    
        CCell
        CRCell
    end
    subgraph Chatbook backend
        CIDQ
        CIDEQ
        CIDNone
        RECO
        CNCO
        CODB
    end
    subgraph Prompt processing
        PDB
        LLMProm
        PromParse
        KPFQ
        PromExp 
    end
    subgraph LLM interaction
      COEval
      LLMFunc
      PaLM
      OpenAI
    end
    CCell --> CIDQ
    CIDQ --> |yes| CIDEQ
    CIDEQ --> |yes| RECO
    RECO --> PromParse
    COEval --> CRCell
    CIDEQ -.- CODB
    CIDEQ --> |no| CNCO
    LLMFunc -.- CNCO -.- CODB
    CNCO --> PromParse --> KPFQ
    KPFQ --> |yes| PromExp
    KPFQ --> |no| COEval
    PromParse -.- LLMProm 
    PromExp -.- LLMProm
    PromExp --> COEval 
    LLMProm -.- PDB
    CIDQ --> |no| CIDNone
    CIDNone --> CIDEQ
    COEval -.- LLMFunc
    LLMFunc <-.-> OpenAI
    LLMFunc <-.-> PaLM

Here is an example of prompt expansion in a generic LLM chat cell and chat meta cell showing the content of the corresponding chat object:


Command Line Interface

Playground access

The package provides a Command Line Interface (CLI) script:

llm-prompt --help
# Usage:
#   llm-prompt <name> [<args> ...] -- Retrieves prompts text for given names or regexes.
#   
#     <name>          Name of a prompt or a regex. (E.g. 'rx/ ^ Em .* /').
#     [<args> ...]    Arguments for the prompt (if applicable).

Here is an example with a prompt name:

llm-prompt NothingElse RAKU
# ONLY give output in the form of a RAKU.
# Never explain, suggest, or converse. Only return output in the specified form.
# If code is requested, give only code, no explanations or accompanying text.
# If a table is requested, give only a table, no other explanations or accompanying text.
# Do not describe your output. 
# Do not explain your output. 
# Do not suggest anything. 
# Do not respond with anything other than the singularly demanded output. 
# Do not apologize if you are incorrect, simply try again, never apologize or add text.
# Do not add anything to the output, give only the output as requested.
Your outputs can take any form as long as requested.

Here is an example with a regex:

llm-prompt 'rx/ ^ N .* /'
# NarrativeToResume => Rewrite narrative text as a resume
# NarrativeToScript => Rewrite a block of prose as a screenplay or stage play
# NerdSpeak => All the nerd, minus the pocket protector
# NothingElse => Give output in specified form, no other additions
# NumericOnly => Modify results to give numerical responses only
# NutritionistBot => Personal nutrition advisor AI

TODO


References

Articles

[AA1] Anton Antonov, "Workflows with LLM functions", (2023), RakuForPrediction at WordPress.

[SW1] Stephen Wolfram, "The New World of LLM Functions: Integrating LLM Technology into the Wolfram Language", (2023), Stephen Wolfram Writings.

[SW2] Stephen Wolfram, "Prompts for Work & Play: Launching the Wolfram Prompt Repository", (2023), Stephen Wolfram Writings.

Packages, paclets, repositories

[AAp1] Anton Antonov, LLM::Prompts Raku package, (2023), GitHub/antononcube.

[AAp2] Anton Antonov, LLM::Functions Raku package, (2023), GitHub/antononcube.

[AAp3] Anton Antonov, Jupyter::Chatbook Raku package, (2023), GitHub/antononcube.

[WRIr1] Wolfram Research, Inc., Wolfram Prompt Repository