Jinbo Wang, Senior Software Engineer, Microsoft Martin Lippert, Spring Engineering, Broadcom October 2024 GitHub Copilot + Spring: Making AI-assisted Coding Spring aware Copyright © 2024 Broadcom, Inc. or its affiliates.

How can AI assist Java development experience in IDE?

The Power of GitHub Copilot Code Completion Explain Fix Generate Docs Generate Tests Chat & Inline Chat

Suggest Refactoring: Use AI to plan refactoring

Refactor Code in a Specific Pattern - Replace ‘+’, ‘StringBuffer’ to ‘StringBuilder’, ‘Text Block’ - Escape String - Dedup repeated code - Migrate to design patterns (Builder, Singleton, etc.) - Extract long method to smaller ones …

Code Migration: Modernize Legacy Code - Text Block for Multiline Strings - Replace Loop with Streams API - Use Enhanced Switch Expressions - Use Pattern Matching for instance of

IDE Refactoring vs AI Refactoring ► IDE Refactoring - Rule-based, Predefined - Limited by IDE features - Requires developers to understand when and how to use it ► AI Refactoring - Trained on huge amounts of examples, easy to scale - More creative, can handle more complex refactoring pattern - Suggest refactoring the developer might not have considered

Influence the AI assistant (towards making it Spring aware)

In general, there are two ways Influence the AI assistant Changing the model itself (e.g. via fine tuning) Feeding additional information when asking (e.g. better prompts/questions, RAG, etc.)

The Copilot case in Visual Studio Code

Coding with an LLM Type your question

Coding with an LLM Type your question Generate a prompt for the LLM

Coding with an LLM Type your question Generate a prompt for the LLM Add context information

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project

How to enhance Copilot?

APIs for Copilot There are APIs in place within VSCode to interact with Copilot: - API to submit prompts to the LLM yourself (direct use of Copilot LLM) - API to enhance Copilot Chat (open ended questions) - Enhance / change the prompt - Specify tasks - Fill in variables / placeholders - Enhance / modify the result (e.g. insert buttons) - Inline Chats: no extension API available yet - Side note: APIs are still evolving rapidly, moving target

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project Predefine the question, no need for the user to type something

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project Add more context to the prompt, but what exactly?

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project Show - and maybe enhance the response with additional information / actions

Coding with an LLM Type your question Generate a prompt for the LLM Add context information Send final prompt LLM Receive Response Show and interpret the response Put code snippets from response into your project Be smart about where to and how to add code to the project

Making it Spring aware

Concrete Actions 1

Concrete Actions - We identify concrete actions where AI can help and provide a meaningful contribution - We craft the prompt with the necessary input, there is no user interaction - Not open-ended, very limited to a specific case Example: - 👍 Explain SpEL expression - 👍 Explain SQL/HQL/JPQL Query - 👍 Explain Pointcut expression - 🤔 CRON: eh, AI is too much for that, simple logic is enough

Latest Spring Tools release

Full Chat Agent 3

Chat Agent - Enhance the prompt (with random information) - Fully open-ended / chat experience - Enhancing the result (buttons + better code merge) - The more structure you ask for, the better you can interpret / merge the result

The Spring agent in action

The Spring agent in action

The Spring agent in action

The Spring agent in action

Challenges

LLM is changing behind the scenes 1

New Language or Framework Versions 2

Size limitations 3

Is it getting better or worse? 4

Questions

Thank you @https://github.com/testforstephen @martinlippert https://www.linkedin.com/in/martin-lippert-9a1909/ © 2024 Broadcom