Baidu Comate Code Knowledge Enhancement 2.0: First Domestic Intelligent Code Assistant with Real-time Retrieval

On May 18, at the iTechClub North China 7th Internet Technology Elite Summit, Baidu's Engineering Efficiency Director Zang Zhi delivered a keynote speech titled "Towards AI-Native R&D Paradigm with Human-Machine Collaboration." He announced Baidu's intelligent code assistant Comate's latest achievement—Comate Code Knowledge Enhancement 2.0, which is China's first intelligent code assistant supporting real-time retrieval, with over 10,000 built-in GitHub hot code repositories.

Comate is a code completion and recommendation tool built on Baidu's Wenxin large model. It uses deep learning and natural language processing to analyze developers' programming intentions and automatically recommend suitable code snippets and library functions.

The three major advantages of Comate Code Knowledge Enhancement 2.0 were detailed in the presentation. First, it has built-in 10,000+ GitHub hot code libraries covering various languages and tech stacks including Spring, Mybatis, FastAPI, React, Transformer, PaddlePaddle, AutoGPT, and Langchain. Second, it supports web-based online retrieval to obtain the latest technical knowledge in real-time, making it China's first intelligent code assistant with real-time retrieval capability. It can understand web page content from URLs and generate code accordingly. Third, given any API link, it can generate calling code and corresponding test cases efficiently.

Comate can now deeply understand local codebases and internal private domain knowledge, including business interface documents, product requirement documents, test case documents, and service deployment documents. According to public data, 27% of Baidu's daily new code is automatically generated by Comate, with an overall adoption rate of 46%. Comate currently ranks first in intelligent assistant ratings on both VSCode and JetBrains plugin markets with scores of 4.5 and 4.4 respectively.