This topic surveys the LLMs and surrounding AI systems used in industrial predictive maintenance as of 2025 — focusing on on‑device inference, retrieval‑augmented workflows, and the data plumbing needed for dependable, auditable operations. Predictive maintenance now combines sensor timeseries, equipment metadata, and maintenance logs to predict failures and guide technicians. Achieving this requires orchestration of ingestion, feature engineering, and retraining; robust connectors to timeseries and document stores; and mechanisms for low‑latency inference and contextual retrieval. Key component classes include on‑device LLM inference for low‑latency alerts at the edge; data pipeline orchestration (Dagster) to schedule and test ETL/ML workflows; data catalog and lineage to track asset metadata and model inputs for compliance; cloud platform integrations and cloud data platforms for scale and long‑term storage; and database connectors to expose operational data to models. Practical tools in this stack: Dagster for building and validating pipelines; MCP servers for database access (Neo4j for asset/relationship graphs, MongoDB/MongoDB Atlas and Neon for document and Postgres storage); Redis MCP for fast searches and cache-backed retrieval; Minima for on‑prem RAG on local manuals and schematics; and mcp-memory-service for hybrid, lock‑free semantic memory across local and cloud stores. Trends in 2025 emphasize hybrid deployments (edge + cloud), stronger data lineage and cataloging to support model governance, and increased use of RAG and semantic memory to make LLM outputs actionable and auditable. The combination of orchestration (Dagster), MCP‑enabled connectors, on‑prem RAG (Minima), and hybrid memory services forms a practical template for predictable, maintainable predictive‑maintenance AI systems.