Design and Implementation of an AI‑Powered Development Assistant for Internal Use

The author, a senior R&D engineer in the infrastructure team, introduces a company‑wide AI development assistant aimed at improving developer productivity across the organization.

Background

The rapid rise of AIGC has produced notable products such as GitHub Copilot (June 2021) and OpenAI ChatGPT (Nov 2022). While Copilot offers intelligent code completion and ChatGPT provides versatile natural‑language assistance, both suffer from limitations: lack of controllable context, single‑scenario output, limited integration with IDEs, and security/privacy concerns.

Design Thinking

Analyzing the shortcomings of Copilot and ChatGPT leads to a set of design goals for the internal assistant:

Embed the assistant directly in the IDE (IDE‑centric).

Provide core intelligent code completion.

Offer ChatGPT‑like Q&A while lowering the prompting barrier.

Enable bidirectional linking between the chat window and source code.

Incorporate company‑specific knowledge for private queries.

Allow the assistant to trigger auxiliary operations (e.g., meeting‑room reservation).

Typical Features

Automatic Context Filling

When a developer selects code and presses Alt+i , the assistant automatically copies the selected snippet and injects related context (function signatures, surrounding code, imports, similar fragments) into the query box, improving the accuracy of generated code.

Manual Context Specification

Users can manually add context (package, class, code fragment) via the right‑click menu, ensuring the model receives the necessary information when automatic assembly is insufficient.

Custom Prompt Templates

Several built‑in Prompt templates are provided to reduce the learning curve of Prompt engineering. Users can also create their own templates; an example template is shown below:

你现在是一名编程专家，请对代码进行解释，要求：
1、分步骤解释
2、避免逐行翻译代码，不重要的细节可忽略，对代码按逻辑进行适当聚合
3、针对代码中的特殊设计进行解释
要解释的代码是：${code}

The placeholder ${code} is replaced with the user‑provided snippet at runtime.

Quick Error Diagnosis

Developers can select an exception stack trace, press Alt+i , and choose the “Diagnose Exception” Prompt. The assistant then analyses the error and suggests fixes, with future plans to apply the fix directly to the source.

Integration with Internal Systems (PMO)

The assistant can query internal tools such as the PMO system via natural‑language commands (e.g., “show my schedule for today”). This eliminates the need to switch browsers and manually navigate multiple internal portals.

Smart Code Completion

Built on open‑source large language models, the assistant offers low‑latency, cost‑effective code completion within the IDEA plugin. A dedicated testing suite, model deployment pipeline, and engineering effort ensure stable performance.

Effectiveness Data

Productivity gains are quantified through two main metrics:

Code Completion: The ratio of lines generated by the assistant to total committed lines, adjusted by the proportion of time developers spend coding (≈20%). Measured impact: ~2% overall R&D efficiency improvement.

General Q&A: Survey‑based estimation using the formula: total daily answers × answer effectiveness × time saved per effective answer. Measured values: answer effectiveness = 0.482, time saved per answer ≈ 4.3 minutes.

Operational Commands: Time saved per command (e.g., fetching PMO schedule) multiplied by usage frequency yields total time saved.

Future Plans

Upcoming development focuses on three directions:

Expanding the knowledge base to include private company knowledge repositories.

Building code‑centric features such as code visualization, navigation via natural language, automated optimization, and exception auto‑repair.

Providing end‑to‑end workflow integration throughout the development lifecycle.

References to recent AI advancements (GPT‑4 Turbo, custom ChatGPT, Assistants API, Copilot Workspace) are listed at the end of the article.

Overall, the assistant demonstrates how AI can be tightly coupled with IDEs to streamline coding, knowledge retrieval, and routine operations, delivering measurable productivity improvements while addressing security and usability concerns.