kantver - Fotolia
Kite launches enterprise-grade code completion tool
Kite's new enterprise version of its code completion tool, Kite Team Server, targets larger development teams to help them write code faster by automating repetitive tasks.
Kite, a provider of an AI-powered coding assistant by the same name, has delivered Kite Team Server, an enterprise version of its code completion tool.
Built for the enterprise, Kite Team Server is a self-hosted machine learning (ML) engine for what's known as four token autocomplete technology.
"The key internal metric we use to assess our ML models is how many 'words' or tokens of code the model can accurately predict ahead in one of our test files," said Adam Smith, CEO and founder of Kite. "The model we use for Kite Pro and Kite Free can predict two tokens ahead on average. Kite Team Server without custom model training can predict three tokens ahead. When Kite Team Server is trained on a relevant codebase, the model can confidently predict four tokens ahead."
There's competition out there
Kite Team Server builds on the free version of the tool, Kite Free, which Smith said has 400,000 developers using it to help them code 18% faster than before.
"Autocomplete functionality is a natural benefit of machine learning, as past behavior makes for an excellent source of data to train the models," said Jason Bloomberg, an analyst at Intellyx in Suffolk, Va. "Kite is but one example. We're seeing similar offerings in low-code workflow tools with 'next best action,' as well as chatbots and similar products."
Indeed, there are others in the AI autocomplete space, including Codota and Tabnine, which Codota acquired in March of last year.
"In a way, our biggest competitors are the autocomplete engines used by most professional software developers today: Kite Free, Tabnine, IntelliCode in VS Code, and the built-in autocomplete in the JetBrains family of IDEs," Smith said. "All of these solutions, including Kite Free, use the limited compute on the user's CPU to generate completions."
Adam SmithCEO, Kite
GPUs give Kite Enterprise an edge
However, Kite Team Server can provide better code completions by tapping the power of GPUs. The Kite Team Server GPU trains personalized ML models based on a company's proprietary codebase. Completions delivered by Kite Team Server guide developers to repeat idioms and patterns from internal code.
Kite Team Server runs on GPU-equipped servers, as opposed to the current version of Kite, which runs on the CPU inside users' laptops. In addition to more compute power, this approach allows for enterprise-grade security because Kite Team Server can run behind a company's firewall.
"The GPU serves a dual purpose," Smith said. "When completions are requested by users, Kite Team Server is able to provide completions at an ultra-low-latency. Occasionally, the GPU is also used for model training runs using the code that company admins provide to Kite Team Server."
Kite Team Server provides completions in all of the most popular programming languages and IDEs, because enterprise teams don't write code in one language or use one editor, Smith said. The product supports 16 languages and 16 IDEs.
Automating repetitive tasks
Kite's goal is to automate away the repetitive parts of writing code so software developers can focus on programming. Coders spend too much time on repetitive tasks such as looking up documentation on the web, fixing simple errors and writing boilerplate code, Smith said.
Similar to Google's Smart Compose for Gmail, Kite uses deep learning to save developers time by automatically completing their code statements. Also, Kite's Intelligent Snippets enable developers to complete multi-token statements without copying and pasting from a web search or existing codebase.
"We believe most developer teams will transition to a self-hosted autocomplete server with custom ML model training over the next few years," Smith said.
Pricing for Kite Team Server is $40 per user, per month, which is $10 more per month than a Kite Pro license.
Market research firm Cognilytica, based in Ellicott City, Md., said the market for machine learning platforms was $23.2 billion in 2019 and will grow to $126.1 billion by 2025, representing a 33.73% compound annual growth rate.