Skip to main content

Prompt Strategy

The AutoDev Prompt Generation Strategy is the core functionality of AutoDev, which can generate optimal code suggestions based on your code context.

AutoDev Prompt Example

Typically, a prompt corresponding to an instruction consists of the following five components:

  • Action type. For example: Code complete, Translate to Kotlin, etc.
  • Language context (combined with specifications). For example: specifications corresponding to Java, Kotlin, Python.
  • Technology stack context (combined with specifications). For example: specifications for Controller, Service, Repository.
  • Relevant context (ClassProvider). For example: current file, current directory, current project, all files in the current project.
  • Code (PsiElement). Current code.

Different languages implement ContextPrompter through their own modules, such as JavaContextPrompter, KotlinContextPrompter, etc.

Prompt Architecture​

Therefore, AutoDev adopts a modular architecture inspired by Intellij Rust and JetBrains AI Assistant, as shown in the following diagram:

AutoDev Prompt Example

Each language module implements the corresponding language context and technology stack context based on abstract interfaces. This requires reading dependency-related information such as gradle, maven, package.json, etc.

Relevant Context​

AutoDev provides the following types of relevant context:

  • Static code analysis approach: Generates corresponding context information by combining import syntax and function input/output.
    • Corresponding implementation class: [JavaContextPrompter]
  • Cosine Similarity calculation for code chunk similarity among the 20 most recently opened files. This is one of the implementation methods used by GitHub Copilot and JetBrains AI Assistant.
    • Corresponding implementation class: [SimilarChunksWithPaths]

AutoDev Similar Chunk

Dual-Prompt for Hidden Details​

In AutoDev, complex prompts are implemented through two separate prompts, as shown below:

abstract class ContextPrompter {
  open fun displayPrompt(): String = ""
  open fun requestPrompt(): String = ""
  ...
}
  • displayPrompt: The prompt displayed to users, e.g., Code complete, Translate to Kotlin.
  • requestPrompt: The prompt sent to AI services, e.g., Code complete:\n${METHOD_INPUT_OUTPUT}\n${SPEC_controller}\n\n${SELECTION}.

Depending on the scenario, certain details (such as related code chunks, input/output) may be hidden in the user-facing prompt.