GitHub Blog
1. How we built an internal data analytics agent
本文描述了一个内部分析AI代理的实践实现，强调了上下文层的关键作用，对考虑类似工具的数据和工程团队直接有用。 (score: 0.90)
GitHub's data and analytics organization built Qubot, an internal Copilot-powered analytics agent that allows employees to ask questions about data models in natural language and get answers within seconds. Qubot consists of three main components: a user interface available through Slack, VS Code, and Copilot CLI; a context layer that enriches Copilot's reasoning with curated documentation; and a query engine that connects to Kusto and Trino. The context layer proved critical for accuracy and speed, and Qubot reduced strain on the data team while enabling self-service analytics across the company.
- Qubot is a Copilot-powered analytics agent that answers natural language queries about GitHub's data warehouse.
- It has three components: user interface (Slack, VS Code, CLI), context layer (federated documentation), and query engine (Kusto and Trino).
- The context layer enriches Copilot's reasoning and improves accuracy and speed by three times.
- An evaluation framework with test cases, automated runs, and stats aggregation ensures quality.
- Qubot reduced queries in data analytics Slack channels and enabled self-service for non-experts.
- Multiple interfaces lower barriers: no configuration needed in Slack, integrated in VS Code and CLI.
- Federated context contribution incentivizes teams to share knowledge in a single tool.