Here is the quick rundown of why we love Kapa AI assistant: - Plain English answers: It’s trained on all our documentation and gives real, readable answers to everyday questions. - Verifiable sources: This was the main selling point. Our concepts are nuanced, so confident but unsourced answers are dangerous. Every AI response links directly to the source page for quick verification. - Comprehensive & auto-updating: It pulls from the docs site, API reference, and GitHub, and automatically keeps that knowledge fresh. - Built-in convenience: The "Ask AI" button sits right next to the standard search bar, giving users immediate answers instead of making them scroll through articles. - Valuable analytics: We get backend data on exactly what our users are struggling with, which is a huge bonus.

My first exposure to kapa.ai was in December 2024. I remember being attracted by answers that contained inline citations to sources. That wasn't a common feature back then. I also liked that most kapa.ai features were available via API, so we could implement the assistant in our internal processes without having to use a chat in a browser. Before that, we had been experimenting at Espressif with building our own documentation assistants, but they didn't gain much internal traction. Then we saw that Silicon Labs had a chatbot powered by kapa.ai. At the time, we weren't sure if it would be a good approach to give our users technical answers generated by AI that were not fully deterministic. But if the answers have citations, they are more credible, and you can check the source yourself. After talking with kapa.ai, we agreed on a test plan to see if the solution met our requirements. During testing, we progressively added more product documentation and validated performance with Q&A from subject matter experts. I was concerned about correctly answering questions on similarly named chip series (e.g., ESP32 vs. ESP32-S3 vs. ESP32-C3) and directing Western users to English documentation, Chinese users to Chinese documentation. The naming concern turned out not to be an issue, and we deployed two separate assistants to handle the language split. The documentation assistants launched in April 2025 and were quickly adopted by users and internal teams, saving an enormous number of hours. Users get instant technical answers, something impossible for humans to provide at scale. Beyond providing a dependable and reliable service for over a year now, I appreciate that kapa.ai doesn't rest on its laurels. They listen to our feedback and continuously improve the platform's functionality with features like Internal Technical Assistant, Coverage Gaps, Top Questions, MCP support, real-time refresh, support for new source types, and more. The measure of the assistant's success is steadily increasing adoption and the many positive comments users leave about the assistants, as if they were human.

The thing that won us over early was answer quality. The assistant is only ever as good as the documentation behind it, and that's exactly what we wanted: it doesn't make things up. When it isn't sure, it tells you, and it points you to the source page so you can check. We threw jailbreak attempts at it during testing and the safeguards held up. It also handles versioning sensibly.