Demystifying Log Detective: How Packit's New AI Analyzes Build Failures
If you've ever scratched your head over a failed Koji build in a dist-git pull request, you're not alone. Starting this month, Packit is rolling out Log Detective — an AI-powered analysis tool that automates the hunt for what went wrong. This Q&A covers everything you need to know about how it works, what it delivers, and who it's designed for.
What is Log Detective and how does it work with Packit?
Log Detective is an intelligent analysis service that examines failing Koji builds triggered by Packit on dist-git pull requests. Instead of you manually hunting through thousands of lines of build logs, Log Detective automatically requests an analysis when a build fails. The service parses all logs and build artifacts using a combination of the BeeAI Framework and a template-mining algorithm called Drain. It extracts only the most relevant snippets, which are then analyzed by a relatively small language model. This focused approach saves tokens and reduces analysis time while still pinpointing the root cause of the failure.
How do I trigger a Log Detective analysis?
You don't have to do anything special. In Packit, a build failure automatically sends a request to the Log Detective interface server. The service handles everything in the background — from selecting which logs to send to tuning the prompt. Once the analysis finishes (usually within seconds to a few minutes), the result appears in the Packit dashboard, linked directly to the pull request that triggered it. There are no extra setup steps, no buttons to click, and no configuration to tweak. Just push your changes and let Log Detective do the detective work.
What makes Log Detective's log parsing so efficient?
Traditional log analysis often involves feeding huge raw log files into an AI model, which is slow and expensive. Log Detective takes a smarter route. Using the Drain template mining algorithm, it extracts small, meaningful snippets from the logs — typically representing just a tiny fraction of the original file size. These snippets are then fed into the language model for analysis. By drastically reducing the amount of data sent to the model, Log Detective saves on tokens and limits analysis time. This efficiency also allows it to work well with smaller models, keeping infrastructure costs low while still producing accurate results.
How does Packit communicate with Log Detective?
The communication flow is carefully designed for reliability and asynchronicity. Packit's service handles the failed Koji build as usual, but now also sends an analysis request to a lightweight interface server dedicated to Log Detective. This server manages all interactions between the two services. Once the analysis is complete, the interface server publishes the results on the Fedora Messaging bus. Packit picks up the results from the bus and displays them in its dashboard, linked to the original pull request. This decoupled architecture ensures that neither service is blocked waiting for the other.
What kind of results does Log Detective provide?
An analysis from Log Detective includes two key pieces of information: a clear statement of what went wrong during the build and, when possible, a suggestion for a solution. For example, it might identify a missing dependency, a syntax error in a spec file, or a conflict with an updated library. Currently, the analysis is based solely on the build logs — Log Detective does not pull in external sources like Git history or package documentation. The result is displayed in the Packit dashboard under the same pull request that triggered the analysis, making it easy to find and act upon.
Who is Log Detective intended to help?
Log Detective is built with newcomers to Fedora package maintenance in mind. If you've been building packages for years, you probably know the common failure patterns by heart. But if you're just starting out, a puzzling build log can be daunting. Log Detective acts as an automated mentor, explaining failures and hinting at fixes. However, because it uses a general-purpose model and lacks access to broader project context, it's not a substitute for experience. It's a tool to lower the barrier to entry, not to replace seasoned maintainers. Experienced users might find its suggestions too basic, but for beginners, it can save hours of frustration.