ChatDBA: An AI‑Powered Database Fault Diagnosis Assistant Using Large Language Models

ChatDBA, developed by Shanghai Aikesheng, is an intelligent assistant for database operations that leverages large language models (LLMs) to provide fault diagnosis, knowledge learning, SQL generation and optimization through conversational interaction.

The early RAG‑based prototype suffered from vague answers, weak logical flow, and model bias, prompting a redesign that introduces a fault‑diagnosis tree, multimodal retrieval, query rewriting, document formatting, and a structured pipeline of environment input processing, model planning, and tool usage.

Key techniques include multi‑path recall (keyword + vector), query rewriting/expansion, multimodal retrieval, vertical domain enhancement, and graph‑RAG. Document processing formats tickets into phenomenon, cause, method, and solution, with scoring and re‑writing loops to ensure completeness.

Memory handling is addressed by summarizing long dialogues and retaining recent turns, while intent recognition uses predefined templates and multi‑intent branching. Observability and evaluation combine long‑chain monitoring with knowledge‑graph‑based explainability.

Future work focuses on multimodal handling, real‑time monitoring integration, knowledge‑graph construction, and personalized recommendation for DBAs.