New AI Helps Engineers Find the Fix When Cloud Services Break

When a big cloud service breaks — your banking app freezes, a video site stalls, an online store won't load — an on-call engineer has to find the right repair instructions fast, usually under heavy pressure. Today that search is mostly manual keyword typing, and past research shows the hunting itself eats up a large share of the total time it takes to fix things. That's the bottleneck this paper attacks.

The team's system, called TSG Suggester, reads a short description of what's going wrong and recommends the matching Troubleshooting Guide (think: the repair manual for that specific failure). Instead of chopping documents into flat text blocks, it turns each guide into a tree that keeps its original section structure, then builds a knowledge graph — a map of how the named parts and problems connect to each other. At search time it blends meaning-based matching with direct name matching.

They tested it on 314 real incidents drawn from 112 guides across 18 service teams at a production incident system. TSG Suggester got the single best answer right 54.78% of the time and included the right guide in its top five 82.48% of the time — beating all four competing methods at every cutoff, and gaining 8.58 points on the closest text-only approach. Two side findings stand out: preserving document structure clearly matters, and, contrary to the authors' own guess, adding captioned screenshots actively hurt, costing 22.64 Top-5 points because generic captions diluted the search quality.

The honest catch: this tool only suggests the guide — a human still diagnoses and does the repair, so it speeds things up rather than replacing anyone. It was also tested inside one company's system with its own writing style, so the same gains may not carry over everywhere. And it needs a well-organized library of guides to begin with. Still, for anyone who has waited on hold while a service was down, shaving minutes off the hunt is real relief.