Fighting logs with Log Detective
summary
A system that diagnoses RPM build failures by combining template mining (using Drain3) with LLMs (local Llama.cpp or hosted models like mistral-0.2), accepting a log link at logdetective.com/explain and returning an explanation in ~1–2 minutes. GitLab integration posts comments on CentOS Stream MRs, a CLI runs entirely locally with quantized models, and the pipeline uses per-snippet annotation then a final synthesis. Over 1,000 human-annotated logs (CDLA‑Permissive‑2.0) exist; accuracy is good for many failures but not perfect and the web annotator is temporarily broken.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 2:27Log Detective in Fedora Copr
- 4:35How it works under the hood
- 10:33GitLab integration feedback
- 13:05CentOS Stream demo
- 15:55Website interface
- 17:18Training data challenges
- 20:36Q&A
- 24:31Q&A
- 28:43Performance evaluation
- 32:37Q&A
- 35:08Closing remarks
related talks
42mFrom Logs to Fixes: Teaching AI to Maintain Software
38mEnhanced Log Analytics Using Generative AI
19mLLM codegen fails and how to stop 'em
37mChallenges and Pitfalls of Applying GenAI in Monitoring
31mAutomating Linux Kernel Documentation for Safety-Criti...
40mInstructLab: Applying Open Source Methods to Building and Trai...
39mCombining the Best of Two Worlds: From TF-IDF to Llama LLM
27mBuilding Useful Apps With Small, Local LLMs