Vulcan: a maintenance assistant that knows every manual and every past repair

Our approach

Manuals, SOPs, P&ID drawings and warranties are split into passages (with OCR for scans), turned into vector embeddings by a local embedding model and stored in PostgreSQL with pgvector. A read-only connector on the plant network syncs the CMMS (direct database read or scheduled export, no cloud service): work orders, parts used, failure codes, run-hours and technician notes. For each question, the agent queries both sources in parallel, filtered on the exact asset: the official repair steps, and how the same fault was fixed on this same machine. An open-source language model (Llama, Qwen or Mistral) running on a GPU server in the plant then writes a short guide citing the manual page and the work-order number, and can draft the new work order with its parts. Speech-to-text (Whisper) also runs locally: no question, document or data ever leaves the site, and the assistant keeps working without internet. Thumbs up / down and a weekly evaluation pipeline flag outdated manuals and parts lists.