AuthoringAssistant

GitHub Actions setup

Ensure the LLM_PROMPT_EXECUTOR_TOKEN secret is set to a personal access token which has been granted the read:packages permission.

Developer setup

Before setting up the project, ensure you have the following dependencies installed:

• Java Development Kit (JDK) 22 (or a more recent version)
• Apache Maven (latest stable version recommended)

Note: Ensure that either Java and Maven are not installed before running this script, or have them preinstalled with Java set to version 22 and JAVA_HOME properly configured.

Automated Installation

To simplify the installation process, you can run the provided script:

./install-dependencies.sh

This script will install the required dependencies automatically.

Manual Configuration

After installation, you need to create a settings.xml file in your Maven configuration directory:

${USER_HOME}/.m2/settings.xml

Copy and paste the following content into the settings.xml file:

<settings xmlns="http://maven.apache.org/SETTINGS/1.0.0"
          xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
          xsi:schemaLocation="http://maven.apache.org/SETTINGS/1.0.0
                      http://maven.apache.org/xsd/settings-1.0.0.xsd">
    <servers>
        <server>
            <id>github</id>
            <username>{GITHUB_USERNAME}</username>
            <password>{GITHUB_TOKEN}</password>
        </server>
    </servers>
</settings>

The {GITHUB_TOKEN} must be set as per LLM_PROMPT_EXECUTOR_TOKEN above.

The Java Application

The folder jar contains the built Java application, which is run through the command:

java -jar jar/prompt-executor.jar agent=string threshold=float systemPromptPath=string settings=string testPath=string numTestToGenerate=int numLearningCaseToGenerate=int numQueryToExecute=int

where:

• agent is the agent class which correspond to the model will be executed
• threshold represents the accuracy threshold value for which the tests will be considered passed (float between [0-1])
• systemPromptPath is the path of the JSON which contains the system-prompt
• settings is the path of the JSON settings file
• testPath is the path of the test cases file
• numTestToGenerate is the number of test cases to generate from a single test case with variables
• numLearningCaseToGenerate is the number of learning cases to generate from a single learning case with variables
• numQueryToExecute is the maximum number of program which will be executed

List of available agents

Class	Model	Token Needed
ClaudeSonnetEvaluatorAgent	claude-sonnet-20241022	yes
GeminiPro15EvaluationAgent	gemini-pro	yes
Gpt4oEvaluationAgent	gpt-4o	yes
Gpt35EvaluationAgent	gpt-3.5	yes
Llama3EvaluatorAgent	llama3:8b	no
Llama31EvaluatorAgent	llama3.1:8b	no

Yarn integration

All the scripts are integrated with yarn scripts.

• yarn ollama-install to install ollama
• yarn pull-llama3 to pull llama3 model
• yarn ollama-serve to start ollama server
• yarn test to launch test with file in input folder
• yarn test-mock to launch test with a mocked agent.

Websites

Paths are relative to folder containing these instructions. $WEBSITE_NAME refers to any of the website folders under website.

Bundling/Testing Websites

To bundle a website:

1. Run yarn fluid publish -w $WEBSITE_NAME -l

This will create a folder in dist with a Lisp-cased version of $WEBSITE_NAME.

To test the website in the browser:

1. run npx http-serve dist/$WEBSITE_NAME_LISP_CASE -a 127.0.0.1 -c-1
2. Open browser at localhost

To run your website tests:

1. Run yarn website-test $WEBSITE_NAME_LISP_CASE