GitHub Copilot Workspace Index and Retrieval Augmented Generation (RAG)

Source: Make chat an expert in your workspace - Visual Studio Code Documentation

GitHub copilot chat can use an index of the set of files you are working with to provide answers.

Search strategy

Once the index for the work space’s files are created, these are suggested strategies when searching for information:

• Determine information needed for answers like past chat history, file structure, current selections

• Search using:

  • Code search engine using index
  • Local semantic search based on meaning for question and keywords
  • File name and content search (grep)
  • Search on language details like function signatures, parameters, symbols, structure

• Narrow results if too large and include only most relevant content in context for prompt

Workspace Chat in VS Code

Source: Make chat an expert in your workspace

In VS Code chat prompts can use #codebase (chat tool) keyword to provide prompt context and start/update indexes. It will adjust automatically according to the size of the project and resources. The @workspace (chat participant) keyword can also be used as a subject matter expert on the the work space, but #codebase tool is recommended because it allows the language model handling the prompt to execute the tool multiple times and other tools if needed.

All files are indexed except ignored ones like .gitignore specified files. File/code, structure like directory and language structure and editor selection and open file is used in context.

Workspace Context in Chat

Workspace context use depends on:

• Agent/plan - agentic codebase search and may have more targeted searches
• Ask - ask may determine intent requires codebase context and use it or explicitly use the #codebase tool
• Edit - similar to ask and after intent detection, relevant code snippets will be added in chat context

See Also

• Retrieval Augmented Generation (RAG) - Retrieval Augmented Generation (RAG)