Locally hosted conversational process mining for production data via graph-based retrieval-augmented generation (GraphRAG)

A secure, on-premises conversational process mining chatbot that enables manufacturing companies to analyze production event logs with natural language queries is presented in this paper. The system adopts a graph-based retrieval-augmented generation (GraphRAG) approach to process mining: PM4Py discovers the process from logs, which is converted into concise activity-, path-, and variant-level facts and stores them with the process graph in a graph database. A hybrid retriever and a lightweight cross-encoder reranker select focused evidence for a compact large language model (LLM), enabling accurate answers about flows, bottlenecks, and variants. A key contribution is the fully local, open-source design covering the embedding model, graph database, reranker, and LLM, to ensure the privacy of sensitive and confidential data. The architecture is detailed using the Active Structure methodology, and the deployment is demonstrated with the Analytics Canvas in a representative use case. The result is a practical, private way for manufacturers to ask questions of their data and act on the insights.