Copilot Enterprise Introduces Search and Customized Best Practices (2024)

GitHub Copilot Enterprise is moving out of beta and into general availability with new features, including the ability to train Copilot on an organization’s best practices and documentation. Copilot also now integrates a Bing search to provide up-to-date context in the chat.

The first feature means organizations can add their own custom best practices and documentation to Copilot to support developers, said Mario Rodriguez, GitHub’s vice president of product who oversees Copilot.

“One of the things that they told us is, ‘Look, GitHub, I have all of these best practices that I want our developers to follow, and I have it sometimes in documentation and I have it sometimes in all these places, and it doesn’t get used,’” Rodriguez said. “Can we just have Copilot have access to those sets of best practices and the developer can ask Copilot, then Copilot will answer it to them.”

The feature, called knowledge bases, essentially is a customization of the model that GitHub can make for an enterprise using text or markdown files from the organization’s repos, Rodriguez explained. It’s a technical process that requires GitHub to perform the actual customization, he added.

Internally, GitHub uses knowledge bases to instill its engineering, security and accessibility best practices into Copilot. It might, for example, be able to tell developers how the organization deploys a Kubernetes cluster, he said.

“Those three repos have a set of best practices we want our developers to follow and understand the why we do things,” he said. “It’s really customized to the way your organization does it and that, by itself, is huge for these enterprises as well.”

Copilot Generates Pull Request Summaries

Copilot can also generate GitHub pull request summaries and analyze pull request (PR) diffs for developers, added Rodriguez.

“You could easily spend two hours summarizing that overall. Wouldn’t it be great for you to say, ‘Hey, Copilot, see all of the changes that are made in those files? Summarize it for me and put it in the description of the PR so the reviewers, when they see it, they could understand what are the changes that I made and why are those changes there,’” he said. “It’s a productivity enhancement to the author of the PR.”

Some pull requests can change 1,000 of files, he added, which can make summarizing all the changes hard and time-consuming for developers.

GitHub also is being expanded to offer summaries of the diff, or the difference between two versions of a file.

“Whenever you are a reviewer you’re like, ‘OK, what changed in this codebase so I could really understand?’ and so, if you just got out of a meeting and you go into a code review, you just get a sense of the map of what the changes happened,” he said. “Now, you could just ask a Copilot and say, ‘Describe to me what this diff is all about. Give me context of what it’s really changing.’ Then I get that context, and then I could provide better code review to my colleague overall.”

That feature was added to the beta in November.

Copilot Going Forward

Future plans include the ability to fine-tune code with an enterprise’s own code. That’s currently in alpha, with about 10 customers participating. There are plans for a beta before moving the feature into general availability.

“By fine-tuning, you’re able to then get the suggestion acceptance rate to go up, increment-like significantly, and then your productivity increases because of that,” he said. “So obscure languages that are not in the normal training set, fine-tuning is an amazing way of doing that.”

One use case might be with how an organization does C++ programming. The implementation of C++ can vary greatly by organization, he said, even though it’s the same language, because the notation and the way the code is implemented may vary.

It also would allow for companies to train Copilot on internal tools and libraries, which would improve the suggestions the code completer offers, he added.
GitHub’s goal is to infuse AI throughout the software development life cycle, he said, and it’s happening fast: In fact, it’s the fastest adoption cycle Rodriguez has seen in his career.

“We’re past the point where AI is hype on this,” Rodriguez told The New Stack. “We have over 50,000 organizations now using Copilot, over 1.3 million paid users. We have crossed the chasm. This is not an early adopter thing anymore. We’re really now in the era of, instead of digital transformation, in our opinion, we’re in the era of AI transformation at this moment.“