Prompt Design - amosproj/amos2025ss04-ai-driven-testing GitHub Wiki

Our prompt consist of 4 parts, the source_code.txt, the user_message.txt, the system message and additional context.

Base case:

• source_code.txt: Contains Python code that we want to work with
• user_message.txt: Holds instructions specifying what actions to perform with the code, like writing unit-tests or editing something
• system message: This sets the role (a helpful assistant) and the output-format (Markdown code)

Process:

1. All these base values are set with the CLI and saved in our promt_data object.
2. Once the Model is loaded, the prompt is efficiently pieced together and the remaining space for the context is filled with information form the include_project module or the internet_search module if they are activated.

Enhancements with Modules

include_project:

1. Embed the source_code.txt and user_message.txt into the RAG
2. Extract closely related IDs
3. Transform their content back into plain text with the embedding model
4. Add this as context

internet_search:

1. Uses KBERT to extract relevant keywords from source_code.txt and user_message.txt
2. Perform targeted web searches
3. Clean findings
4. Add this as context