What documents can be indexed?

Any PDF, DOCX, or TXT document can be ingested: standard operating procedures, work instructions, maintenance manuals, safety data sheets, incident reports, and quality control guides. Documents are chunked, embedded, and stored in the local ChromaDB vector database.

AI Manufacturing Diagnostics — On-Premise AI Copilot for Shopfloors

What is RAG-Based Manufacturing Diagnostics?

RAG (Retrieval-Augmented Generation) diagnostics is an AI approach where the system answers questions by retrieving relevant content from your documents first, then generating an answer grounded in those documents. Unlike a general-purpose LLM, which answers from training data alone, a RAG system searches your specific knowledge base — SOPs, work instructions, maintenance records, safety data sheets — before generating any response.

The result: AI answers that reference your procedures, not generic internet knowledge. An operator asking "Why is line A02 showing reduced cycle time on station ST05?" receives a checklist derived from the maintenance manual for that specific machine type, not a generic suggestion.

How Shopfloor Copilot AI Diagnostics Works

The diagnostic pipeline runs in six steps, entirely on your server:

1. Document ingestion — PDFs, DOCXs, and text files are chunked and embedded into ChromaDB (the local vector store)
2. Signal context — The system reads live OEE data, active alerts, and downtime events for the station in question
3. Semantic retrieval — ChromaDB retrieves the most relevant document chunks for the fault or question
4. Context assembly — Retrieved documents + live machine context are assembled into a structured prompt
5. Local LLM inference — Ollama runs the language model inference on-premise (Llama 3, Mistral, or any compatible model)
6. Cited response — The operator receives a step-by-step checklist with source citations showing which document each recommendation came from

🧠 AI Root Cause Copilot

Analyses current station OEE, active alerts, and downtime history. Proposes ranked root cause hypotheses with supporting evidence from your SOPs.

💬 Operator Q&A Assistant

Natural language Q&A interface. Ask "What is the torque spec for station ST05 assembly?" and get a cited answer from your work instructions.

📄 Document Knowledge Base

Ingest an unlimited number of documents. PDFs, DOCX, TXT supported. ChromaDB stores and indexes all content locally.

🔒 100% On-Premise LLM

Powered by Ollama running locally. No API key needed. No production data leaves your server. Compatible with Llama 3, Mistral, Gemma, and others.

📊 Evidence Citations

Every AI answer includes source citations: document name, relevant excerpt, and similarity score. Operators can verify the source before acting.

🛡️ Hallucination Prevention

Answers are restricted to retrieved document content. If no relevant document exists, the system says so — it does not fabricate procedures.

Why On-Premise AI Matters for Manufacturing

Most AI diagnostic tools rely on cloud LLMs (GPT-4, Claude, Gemini). This means your production data — machine fault codes, OEE values, operator notes — is transmitted to external servers. For manufacturers in automotive (IATF 16949), aerospace (AS9100D), pharmaceutical (GxP), and defence (ITAR) sectors, this is a compliance and security problem.

Shopfloor Copilot solves this with a local-first AI architecture:

Ollama runs the LLM model on your server's CPU/GPU — no internet connection required
ChromaDB stores all vector embeddings locally — no external vector database API
All inference runs within your OT network perimeter
Compatible with air-gapped environments

What Documents Can Be Ingested?

Any document in your existing quality management system (QMS) or maintenance system can be indexed:

Standard Operating Procedures (SOPs)
Work Instructions (WIs)
Maintenance Manuals and OEM documentation
Safety Data Sheets (SDS)
Non-conformance reports (NCRs) and corrective action records
Equipment calibration records
Historical incident reports

Supported file formats: PDF, DOCX, TXT, Markdown. Files are ingested via the platform's document management interface.

Frequently Asked Questions

What is RAG-based manufacturing diagnostics?

RAG (Retrieval-Augmented Generation) diagnostics means the AI answers are grounded in your actual documents — SOPs, work instructions, maintenance manuals. When a machine fault occurs, the AI searches your document knowledge base for relevant procedures, then generates a step-by-step checklist grounded in those real documents. This eliminates AI hallucinations.

Does the AI send any data to the cloud?

No. Shopfloor Copilot runs local LLM inference via Ollama on your own server. No data leaves your facility. No API calls to OpenAI, Anthropic, or any external service. Suitable for air-gapped environments and OT security requirements.

Which LLM models are supported?

Any model supported by Ollama: Llama 3 (8B, 70B), Mistral, Mixtral, Gemma, Phi-3, and others. The default configuration uses Llama 3 8B, which runs well on a server with 32GB RAM without a dedicated GPU.

See AI Diagnostics in Action

Request a live demo showing real-time fault analysis, operator Q&A with document citations, and local LLM inference — all running within a simulated factory environment.

Explore Platform →