All projects

shopfloor

An enterprise-oriented AI knowledge assistant designed for industrial environments where technical documentation is extensive, sensitive, and frequently updated. The system processes equipment manuals, welding procedures, operating instructions, and spare-parts catalogs, allowing operators to retrieve accurate information through natural-language questions. Built with Python, FastAPI, PostgreSQL, pgvector, local embeddings, and self-hosted language models, the entire platform operates without internet access. Its retrieval architecture verifies numerical information against original documents and attaches source references, including document revisions and page numbers. When sufficient evidence is unavailable, the system explicitly identifies the information gap rather than generating unsupported answers. Quality and retrieval accuracy are continuously evaluated through Groundcheck, using a dedicated 39-question benchmark.

Python · FastAPI · pgvector · Local LLMs · RAG · Docling