RAG SYSTEM APPLICATION BASED ON LLM FOR DECISION SUPPORT DURING DISASTER RECOVERY IN CLOUDS

Abstract

When cloud infrastructure fails, on-call engineers face a challenge: making expert-level decisions under severe time pressure, often with incomplete information and tools that were not designed for the pace of modern incident response. Static runbooks stored in wiki systems and ITSM platforms quickly become outdated, rely on keyword-based searches that assume familiarity with exact terminology, and offer no mechanism for synthesizing guidance across multiple knowledge sources in real time.This article proposes and experimentally validates a decision-support system for AWS cloud disaster recovery based on Retrieval-Augmented Generation (RAG)–a paradigm that combines dense vector retrieval with the generative capabilities of large language models. The system operates in two phases. In the knowledge ingestion phase, an Anthropic Claude Sonnet model, accessed via AWS Bedrock, automatically generates structured recovery runbooks for 75 incident scenarios across 10 infrastructure categories, each following a consistent eight-section format that covers symptoms, immediate mitigation, investigation steps, recovery procedure, verification, and prevention. In the query phase, an engineer's natural-language incident description is embedded with Amazon Titan Embeddings v2 and matched against the stored knowledge base via pgvector cosine-similarity search, after which the retrieved context is used to generate a structured, actionable response in approximately 3 seconds. Several quality determination tests were performed, and conclusions were drawn, as well as the next optimization steps.The proposed approach reduces manual overhead in knowledge base maintenance, improves retrieval precision via optional category filtering, and demonstrates robustness in scenarios where no exact runbook match exists