dash_agent 0.4.0

Framework package to create and publish dash agents for commanddash

Dash Agent

Dash Agent is a framework enabling you to create and publish agents inside CommandDash marketplace.

Getting started

You can create a new dash agent using dash-cli which automatically creates a starter project for you.

dart pub global activate dash-cli

After activating the CLI, run the command to create your agent project.

Replace {{agent}} with the unique name of your agent.

dash_cli create {{agent}}

Usage

This package contains the building blocks for creating dash agents:

AgentConfiguration

The main part of the framework that glues together your agent configuration is AgentConfiguration.

Sample example of AgentConfiguration:

class MyAgent extends AgentConfiguration {
  final docsSource = DocsDataSource();
  final blogsSource = BlogsDataSource();

  @override
  List<DataSource> get registerDataSources => [docsSource, blogsSource];

  @override
  List<Command> get registerSupportedCommands =>
      [AskCommand(docsSource: docsSource)];

  @override
  Metadata get metadata => Metadata(
    name: 'My Agent',
    avatarProfile: 'assets/images/agent_avatar.png',
    tags: ['flutter', 'dart'],
  );

  @override
  String get registerSystemPrompt => '''You are a Flutter expert who answers user queries related to the framework.

  Note:
  1. If the references don't address the question, state that "I couldn't fetch your answer from the doc sources, but I'll try to answer from my own knowledge".
  2. Be truthful, complete and detailed with your responses and include code snippets wherever required''';
}

The above AgentConfiguration registers your data sources, supported commands, metadata, and system prompt.

• DataSource: Data sources are crucial for providing your agent with the information it needs. This could be anything from official documentation, blog posts, code examples, or even custom knowledge bases. In the example above, docsSource and blogsSource represent potential data sources.
• Command: Commands define the specialized actions your agent can perform. These are like building blocks for more complex tasks. The example uses the AskCommand which allows your agent to answer user queries based on the provided docsSource.
• Metadata: Metadata provides information about your agent, such as its name, display avatar, and associated tags. This helps users understand what the agent is about and how it can be used.
• System Prompt: The system prompt sets the initial context for your agent. This is like giving it instructions or defining its persona. In this example, the prompt makes it clear that "My Agent" is a Flutter expert, how it should handle responses, and the desired level of detail in its answers.

Datasource

As described above data sources let you attach the data that your agent will need to perform its tasks. It can be anything from raw texts, JSON, file data, or webpages or even repos.

Example for DataSource:

class DocsDataSource extends DataSource {

 /// Enables you to provide data stored in files and directories in your 
 /// local system.
 @override
 List<FileDataObject> get fileObjects => [
       FileDataObject.fromFile(File(
           'your_file_path')),
        FileDataObject.fromDirectory(Directory(
            'directory_path_to_data_source'))
     ];
 

 /// Enables you to provide raw string and json data
 @override
 List<ProjectDataObject> get projectObjects =>
     [ProjectDataObject.fromText('Data in form of raw text')];


 /// Enables your agent to use web pages by indexing specified web page 
 /// URLs or sitemaps or even github repos in the object.
  @override
 List<WebDataObject> get webObjects =>
     [WebDataObject.fromWebPage('https://sampleurl.com'),
     WebDataObject.fromWebPage('https://sampleurl.com', deepCrawl: true), // to index all the pages of sampleurl.com
     WebDataObject.fromSiteMap('https://sampleurl.com/sitemap.xml'),
     WebDataObject.fromGithub('https://github.com/user/repo', '<personal access token>', codeFilter: CodeFilter(pathRegex: '.*\.dart'),  issueFilter: IssueFilter(labels: ['bug']))];
}

Note: At the moment, storing only text-based files like code, markdown or raw text is supported.

Commands

Commands are the specialised tasks you want your agents to perform (such as refactoring, code generation, code analysis, etc). Once the agent is published, users can invoke this command in the command dash client (such as VS Code extensions) and use it.

Sample example for Command object is shared below:

class AskCommand extends Command {
  AskCommand({required this.docsSource});

  final DataSource docsSource;

  // Inputs
  final userQuery = StringInput('Your query');
  final codeAttachment = CodeInput('Code Attachment');
  
  /// Unique identifier of the command
  @override
  String get slug => '/ask';

  /// Brief description of the command
  @override
  String get intent => 'Ask me anything';

  /// List of `DashInput`s that will be used in the command in its
  /// lifecycle
  @override
  List<DashInput> get registerInputs => [userQuery, codeAttachment];

  /// Series of operations that need to be performed for a command to
  /// finish its task
  @override
  List<Step> get steps {
      // Outputs
     final matchingDocuments = MatchDocumentObject();
     final queryOutput = QueryOutput();

      return [
        MatchingDocumentStep(
            query: '$userQuery$codeAttachment',
            dataSources: [docsSource],
            output: matchingDocuments),
        PromptQueryStep(
            prompt:
                '''You are an X agent. 
                
                Here is the $userQuery, here is the document references: $matchingDocuments. 
                
                Answer the user's query.''',
            postProcessKind: PostProcessKind.raw,
            output: queryOutput),
        AppendToChatStep(
            value:
                'This was your query: $userQuery and here is your output: $queryOutput'),
      ];
}
 
  /// Phrase that will be shown to user when the command is invoked
  @override
  String get textFieldLayout =>
      "Hi, I'm here to help you. $userQuery $codeAttachment";
}

One of the important elements of the Command object is the series of steps that you will be passing that will help the command to execute its tasks by performing the mini-tasks required to be performed by the main task.

Step

Currently supported steps that are available for you to leverage are shared below:

• MatchDocumentStep - Helps you find the matching document from the provided data source form DataSource objects.
• WorkspaceQueryStep - Helps you find the matching code snippets from the user's project.
• PromptQueryStep - Enables you to perform a request to the LLM model with your customised prompt and instruction from the user to perform get generated code or any other general response that can be either used for the next steps or passed back to the user as the final response.
• AppendToChatStep - Enables you to append the response (anything like code, feedback, or general response) to the command dash client chat box.

In future more steps will be included in the list as the framework evolves.

Testing Your Agents

After deploying your agent for testing using the dash_cli publish --test command, you'll find your agent labeled as test in the CommandDash agent marketplace, as illustrated below:

Simply click on the "install" button, and you'll be able to test your newly created agent within the extension.

Note: Access the CommandDash agent marketplace page by clicking on the marketplace icon (highlighted in the red square) within the CommandDash extension:

Additional information

We welcome the Flutter and AI enthusiasts likewise to contribute to this amazing open-source framework. You can contribute in the following ways:

• File feature requests: Suggest features that'll make your development process easier in the issues board.

• Pick up open issues: Pick up and fix existing issues open to the community in issues board.

• Participate in discussions: Help by sharing your ideas in the active discussions in our community slack.

Community

Connect with like-minded people building with Flutter and using AI to do so, every step of the way :D Join Now