🏭 IndusAI — Industrial AI Assistant

v1.0 · PLC Fault Analysis with RAG + Local LLM

---

🎯 What It Does

IndusAI is a privacy-first industrial AI assistant that runs 100% locally. It analyzes PLC (Programmable Logic Controller) fault logs using:

• Deterministic statistics — occurrence frequency, burst detection, co-occurrence analysis
• RAG (Retrieval-Augmented Generation) — retrieves relevant sections from indexed industrial documentation
• Local LLM reasoning — Ollama + Mistral 7B generates structured fault explanations (no cloud APIs)
• Strict output validation — hallucinated PLC tags are detected and stripped before the response reaches the user

✨ Features

Feature	Details
📂 CSV Upload	Drop any PLC fault log — 20+ column aliases auto-mapped
🔍 RAG Retrieval	Hybrid vector + keyword search across indexed documentation
🤖 Local LLM	Mistral 7B via Ollama — fully offline, no OpenAI
📊 Deterministic Stats	Occurrences, burst detection, co-faults — never recomputed by LLM
❓ Custom Q&A	Ask a specific question about any fault row
🏥 System Health	Real-time header badge — LLM, RAG, Vector DB status
🚫 No Silent Fallback	Explicit error if LLM/RAG unreachable — no misleading mock responses

🏗️ Architecture

CSV Upload ──→ Schema Normalizer ──→ FaultService (in-memory)
                                          │
User clicks "Ask AI" ──────────────→ FaultAnalysisOrchestrator
                                          ├── SystemHealthService  (pre-flight check)
                                          ├── FaultService         (deterministic stats)
                                          ├── RAGService           (HybridRetriever → Qdrant)
                                          ├── OllamaLLM            (Mistral 7B, local)
                                          └── FaultResponseValidator (hallucination check)
                                                    │
                                            FaultAnalysisV2Response ──→ React UI

🚀 Quick Start

Prerequisites

• Python 3.11+
• Node.js 18+
• Ollama — brew install ollama
• Qdrant binary (see below)

1. Backend

cd industrial_ai_assistant

# Create virtualenv
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Create data dir
mkdir -p data

# Start Qdrant (download binary)
curl -fsSL "https://github.com/qdrant/qdrant/releases/download/v1.11.5/qdrant-aarch64-apple-darwin.tar.gz" \
  -o /tmp/qdrant.tar.gz && tar -xzf /tmp/qdrant.tar.gz -C /tmp
/tmp/qdrant &

# Start Ollama + pull model
ollama serve &
ollama pull mistral

# Start backend
PYTHONPATH=. ./venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8001 --reload

2. Frontend

cd industrial_ai_assistant/client
npm install
npm run dev
# → http://localhost:5173

3. Verify everything is healthy

curl http://localhost:8001/api/system/health
# {"status":"healthy","llm_connected":true,"rag_connected":true,"vector_store_connected":true}

📡 API Endpoints

Method	Path	Description
POST	/api/fault/upload	Upload PLC fault CSV
GET	/api/fault/list	Paginated fault rows
GET	/api/fault/summary	Aggregate statistics
GET	/api/fault/detail	Single row + history
POST	/api/fault/analyze	LLM analysis (+ optional question)
GET	/api/system/health	System connectivity status
POST	/api/chat	RAG-based doc Q&A

🧪 Tests

cd industrial_ai_assistant
PYTHONPATH=. ./venv/bin/pytest tests/ -v
# 47 tests — schema normalization, upload, summary, analysis, RAG integration

📁 Project Structure

industrial_ai_assistant/
├── app/
│   ├── api/             # FastAPI routes
│   ├── core/            # Interfaces, exceptions, schemas
│   ├── config/          # DI container, settings
│   ├── embeddings/      # SentenceTransformer + mock
│   ├── llm/             # Ollama + mock LLM
│   ├── models/          # Pydantic v2 models
│   ├── retrieval/       # HybridRetriever, keyword search
│   ├── services/        # Orchestrators, health, RAG, validation
│   ├── utils/           # Schema normalizer, stats, confidence
│   └── vector_store/    # Qdrant + in-memory store
├── client/              # React 18 + Tailwind frontend
│   └── src/
│       ├── api/         # faultApi.js, systemApi.js
│       ├── components/  # Header, Dashboard panels
│       └── pages/       # LogsPage, ChatPage
├── tests/               # pytest test suite (47 tests)
└── docs/                # Evaluation checklist, risk analysis

⚙️ Configuration

Copy .env.example to .env and adjust:

LLM_PROVIDER=ollama          # or mock
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=mistral
VECTOR_STORE_TYPE=qdrant     # or in_memory
QDRANT_HOST=localhost
QDRANT_PORT=6333
EMBEDDING_PROVIDER=sentence_transformers
DEBUG_LLM=false              # Set true to log full LLM prompts

🔒 Privacy

All processing is 100% local:

• LLM runs via Ollama on your machine
• Documents indexed in local Qdrant instance
• No data leaves your network

📄 License

MIT — see LICENSE

---

Built with ❤️ for industrial engineers who need reliable, explainable AI — not black boxes.