/prompt_manager

Provides a simple way to manage prompts for a generative AI system.

Primary LanguageRubyMIT LicenseMIT

PromptManager

Manage the parameterized prompts (text) used in generative AI (aka chatGPT, OpenAI, et.al.) using storage adapters such as FileSystemAdapter, SqliteAdapter and ActiveRecordAdapter.

Breaking Change in version 0.3.0 - The value of the parameters Hash for a keyword is now an Array instead of a single value. The last value in the Array is always the most recent value used for the given keyword. This was done to support the use of a Readline::History object editing in the aia CLI tool

Table of Contents

Installation

Install the gem and add to the application's Gemfile by executing:

bundle add prompt_manager

If bundler is not being used to manage dependencies, install the gem by executing:

gem install prompt_manager

Usage

See examples/simple.rb

See also examples/using_search_proc.rb

Overview

The prompt_manager gem provides functionality to manage prompts that have keywords and directives for use with generative AI processes.

Generative AI (gen-AI)

Gen-AI deals with the conversion (some would say execution) of a human natural language text (the "prompt") into somthing else using what are known as large language models (LLM) such as those available from OpenAI. A parameterized prompt is one in which there are embedded keywords (parameters) which are place holders for other text to be inserted into the prompt.

The prompt_manager uses a regular expression to identify these keywords within the prompt. It uses the keywords as keys in a parameters Hash which is stored with the prompt text in a serialized form - for example as JSON.

What does a keyword look like?

By default, any text matching [UPPERCASE_TEXT] enclosed in square brackets is treated as a keyword. [KEYWORDS CAN ALSO HAVE SPACES] as well as the underscore character.

You can customize the keyword pattern by setting a different regular expression:

# Use {{param}} style instead of [PARAM]
PromptManager::Prompt.parameter_regex = /(\{\{[A-Za-z_]+\}\})/

The regex must include capturing parentheses () to extract the keyword. The default regex is /(\[[A-Z _|]+\])/.

All about directives

A directive is a line in the prompt text that starts with the two characters '//' - slash slash - just like in the old days of IBM JCL - Job Control Language. A prompt can have zero or more directives. Directives can have parameters and can make use of keywords.

The prompt_manager only collects directives. It extracts keywords from directive lines and provides the substitution of those keywords with other text just like it does for the prompt.

Example Prompt with Directives

Here is an example prompt text file with comments, directives and keywords:

# prompts/sing_a_song.txt
# Desc: Has the computer sing a song

//TextToSpeech [LANGUAGE] [VOICE NAME]

Say the lyrics to the song [SONG NAME].  Please provide only the lyrics without commentary.

__END__
Computers will never replace Frank Sinatra
Accessing Directives

Getting directives from a prompt is as easy as getting the kewyords:

prompt = PromptManager::Prompt.new(...)
prompt.keywords   #=> an Array
prompt.directives #=> an Array of entries like: ['directive', 'parameters']

# to_s builds the prompt by substituting
# values for keywords amd removing comments.
# The resulting text contains directives and
# prompt text ready for the LLM process.
puts prompt.to_s  

The entries in the Array returned by the prompt.directives method is in the order that the directives were defined within the prompt. Each entry has two elements:

  • directive name (without the // characters)
  • parameter string for the directive
Dynamic Directives

Since directies are collected after the keywords in the prompt have been substituted for their values, it is possible to have dynamically generated directives as part of a prompt. For example:

//[COMMAND] [OPTIONS]
# or
[SOMETHING]

... where [COMMAND] gets replaced by some directive name. [SOMETHING] could be replaced by "//directive options"

Executing Directives

The prompt_manager gem only collects directives. Executing those directives is left up to some down stream process. Here are some ideas on how directives could be used in prompt downstream process:

  • "//model gpt-5" could be used to set the LLM model to be used for a specific prompt.
  • "//backend mods" could be used to set the backend prompt processor on the command line to be the mods utility.
  • "//include path_to_file" could be used to add the contents of a file to the prompt.
  • "//chat" could be used to send the prompts and then start up a chat session about the prompt and its response.

Its all up to how your application wants to support directives or not.

Comments Are Ignored

The prompt_manager gem ignores comments. A line that begins with the '#' - pound (aka hash) character - is a line comment. Any lines that follow a line that is 'END at the end of a file are considered comments. Basically the 'END' the end of the file. Nothing is process following that line.

The gem also ignores blank lines.

Storage Adapters

A storage adapter is a class instance that ties the PromptManager::Prompt class to a storage facility that holds the actual prompts. Currently there are 3 storage adapters planned for implementation.

The PromptManager::Prompt to support a small set of methods. A storage adapter can provide "extra" class or instance methods that can be used through the Prompt class. See the test/prompt_manager/prompt_test.rb for guidance on creating a new storage adapter.

FileSystemAdapter

This is the first storage adapter developed. It saves prompts as text files within the file system inside a designated prompts_dir (directory) such as ~/.prompts or where it makes the most sense to you. Another example would be to have your directory on a shared file system so that others can use the same prompts.

The prompt ID is the basename of the text file. For example todo.txt is the file for the prompt ID todo (see the examples directory.)

The parameters for the todo prompt ID are saved in the same directory as todo.txt in a JSON file named todo.json (also in the examples directory.)

Configuration

Use a config block to establish the configuration for the class.

PromptManager::Storage::FileSystemAdapter.config do |o|
  o.prompts_dir       = "path/to/prompts_directory" 
  o.search_proc       = nil     # default 
  o.prompt_extension  = '.txt'  # default
  o.params_extension  = '.json' # default
end

The config block returns self so that means you can do this to setup the storage adapter with the Prompt class:

PromptManager::Prompt
  .storage_adapter = 
    PromptManager::Storage::FileSystemAdapter
      .config do |config|
        config.prompts_dir = 'path/to/prompts_dir'
      end.new
prompts_dir

This is either a String or a Pathname object. All file paths are maintained in the class as Pathname objects. If you provide a String it will be converted. Relative paths will be converted to absolute paths.

An ArgumentError will be raised when prompts_dir does not exist or if it is not a directory.

search_proc

The default for search_proc is nil which means that the search will be preformed by a default search method which is basically reading all the prompt files to see which ones contain the search term. It will return an Array of prompt IDs for each prompt file found that contains the search term. Its up to the application to select which returned prompt ID to use.

There are faster ways to search and select files. For example there are specialized search and selection utilities that are available for the command line. The examples directory contains a bash script named rgfzf that uses rg (aka ripgrep) to do the searching and fzf to do the selecting.

See examples/using_search_proc.rb

File Extensions

These two configuration options are String objects that must start with a period "." utherwise an ArgumentError will be raised.

  • prompt_extension - default: '.txt'
  • params_extension - default: '.json'

Currently the FileSystemAdapter only supports a JSON serializer for its parameters Hash. Using any other values for these extensions will cause problems.

They exist so that there is a platform on to which other storage adapters can be built or serializers added. This is not currently on the roadmap.

Example Prompt Text File

# ~/.prompts/joke.txt
# Desc: Tell some jokes

Tell me a few [KIND] jokes about [SUBJECT]

Note the command lines at the top. This is a convention I use. It is not part of the software. I find it helpful in documenting the prompt.

Example Prompt Parameters JSON File

{
  "[KIND]": [
    "pun",
    "family friendly"
  ],
  "[SUBJECT]": [
    "parrot",
    "garbage man",
    "snowman",
    "weather girl"
  ]
}

The last value in the keyword's Array is the most recent value used for that keyword. This is a functionality established since v0.3.0. Its purpose is to provide a history of values from which a user can select to repeat a previous value or to select ta previous value and edit it into something new.

Extra Functionality

The FileSystemAdapter adds two new methods for use by the Prompt class:

  • list - returns an Array of prompt IDs
  • path and path(prompt_id) - returns a Pathname object to the prompt file

Use the path(prompt_id) form against the Prompt class Use prompt.path when you have an instance of a Prompt

ActiveRecordAdapter

The ActiveRecordAdapter assumes that there is a database already configured by the application program that is requiring prompt_manager which has a model that contains prompt content. This model must have at least three columns which contain content for:

  • a prompt ID
  • prompt text
  • prompt parameters

The model and the columns for these three elements can have any name. Those names are provided to the ActiveRecordAdapter in its config block.

Configuration

Use a config block to establish the configuration for the class.

The PromptManager::Prompt class expects an instance of a storage adapter class. By convention storage adapter class config methods will return self so that a simple new after the config will establish the instance.

PromptManager::Prompt
  .storage_adapter = 
    PromptManager::Storage::ActiveRecordAdapter.config do |config|
      config.model              = DbPromptModel # any ActiveRecord::Base model
      config.id_column          = :prompt_name
      config.text_column        = :prompt_text
      config.parameters_column  = :prompt_params
    end.new # adapters an instances of the adapter class
model

The model configuration parameter is the actual class name of the ActiveRecord::Base or ApplicationRecord (if you are using a rails application) that contains the content used for prompts.

id_column

The id_column contains the name of the column that contains the "prompt ID" content. It can be either a String or Symbol value.

text_column

The text_column contains name of the column that contains the actual raw text of the prompt. This raw text can include the keywords which will be replaced by values from the parameters Hash. The column name value can be either a String or a Symbol.

parameters_column

The parameters_column contains the name of the column that contains the parameters used to replace keywords in the prompt text. This column in the database model is expected to be serialized. The ActiveRecordAdapter currently has a kludge bit of code that assumes that the serialization is done with JSON. The value of the parameters_column can be either a String or a Symbol.

TODO: fix the kludge so that any serialization can be used.

Other Potential Storage Adapters

There are many possibilities to example this plugin concept of the storage adapter. Here are some for consideration:

  • RedisAdapter - Not sure; isn't redis more temporary oriented?
  • ApiAdapter - use some end-point to CRUD a prompt

Development

Looking for feedback and contributors to enhance the capability of prompt_manager.

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/MadBomber/prompt_manager.

License

The gem is available as open source under the terms of the MIT License.