TACET DISCORD 🧾

Shift-Handoff Copilot & Factory Incident Intelligence Framework
Preserving Senior Technician Tribal Knowledge with Active Evidence Gathering, Local Gemma 4 Reasoning, and Multi-Agent Hallucination Verification.

---

🌟 Overview

TACET DISCORD is an AI-powered industrial incident intelligence framework designed for manufacturing environments. It bridges the gap between senior technician expertise and junior operational personnel by combining visual anomaly detection, telemetry degradation modeling, vector-based semantic incident memory retrieval, and hallucination-verified LLM troubleshooting generation.

---

🔥 Key Features & Capabilities

1. 📊 AI4I 2020 Real Industrial Dataset Seeding

• Real Predictive Maintenance Data: Loaded from 10,000 operational records sampled from the Kaggle/UCI AI4I 2020 dataset.
• Categorized Failure Modes: Supports Tool Wear Failure (TWF), Heat Dissipation Failure (HDF), Power Failure (PWF), Overstrain Failure (OSF), and Random Failures (RNF).
• Telemetry-Rich Schemas: Tracks air/process temperatures, rotational speeds (RPM), torque (Nm), and tool wear times.

2. 🧠 Semantic Text Matcher (sentence-transformers)

• Model: all-MiniLM-L6-v2 generating 384-dimensional dense vector embeddings.
• Paraphrased Query Retrieval: Finds relevant historical incident records even when junior technicians use completely different wording or natural phrasing.
• Dynamic Senior Record Embedding: Automatically computes and indexes semantic vectors whenever a senior technician uploads new data via /records/add.

3. 🤖 Local Gemma 4 & Multi-API Reasoning Engine

• Default Local Model: Runs local Gemma 4 (gemma4:latest) via Ollama (http://localhost:11434) out of the box with zero external API key requirements.
• Hosted API Overrides: Supports automatic fallback/override for hosted providers (GEMINI_API_KEY, GROQ_API_KEY, NVIDIA_API_KEY, OPENAI_API_KEY).
• General Domain Knowledge Fallback: When no strong database records match a query, the system generates best-effort technical troubleshooting steps using general engineering knowledge, explicitly wrapped with a Tier 3 Warning Label (⚠️ General knowledge estimate — confirm with a senior technician).

4. 🛡️ Hallucination Verification Gate

• Claim Extraction & Grounding: Verifies generated LLM diagnosis claims, numerical parameters, and fix action steps against retrieved evidence.
• False Citation Detection: Catches and flags invented database record IDs (🛑 FAILED (Ungrounded Citation)).
• Reasoning Trace Preservation: Records raw LLM outputs, evidence text, and verification claim scores in the reasoning_trace for full inspection in the UI.

5. 🤖 5-Agent Architecture

1. Perception Agent: Feature distribution fitting and visual anomaly heatmap generation (ResNet-18).
2. Correlation Agent: NASA IMS bearing telemetry signal processing & Remaining Useful Life (RUL) estimation.
3. Memory Agent: Hybrid vision embedding + semantic text vector retrieval.
4. Verifier Agent: Cross-agent evidence aggregation and Hallucination Gate validation.
5. Coordinator Agent: Multi-agent workflow orchestration and API response synthesis.

---

🏗️ System Architecture

flowchart TD
    SubGraph1[Junior / Senior Technician] --> |Text Query / Defect Photo| FastAPI[FastAPI Main Server]
    FastAPI --> Coordinator[Coordinator Agent]
    
    subgraph MultiAgentSystem [5-Agent Intelligence Pipeline]
        Coordinator --> Perception[Perception Agent\nResNet-18 Anomaly Heatmap]
        Coordinator --> Correlation[Correlation Agent\nNASA IMS Telemetry RUL]
        Coordinator --> Memory[Memory Agent\nSentenceTransformer all-MiniLM-L6-v2]
        
        Perception --> Verifier[Verifier Agent\nHallucination Check Gate]
        Correlation --> Verifier
        Memory --> Verifier
    end

    Coordinator --> LLMEngine[Grounded LLM Reasoning Engine]
    LLMEngine -->|Local Ollama| Gemma4[Gemma 4: 8.0B]
    LLMEngine -->|Hosted Override| HostedAPIs[Gemini / Groq / NVIDIA / OpenAI]
    
    Verifier --> Output[Verified Copilot Answer + Pipeline Inspection Trace]

Loading

---

🚀 Quickstart Guide

1. Prerequisites

• Python: 3.10+
• Ollama (optional for local Gemma 4 execution): Download Ollama

2. Installation

Clone the repository and install the required dependencies:

git clone https://github.com/developerHarish2007/Tacet-Discord.git
cd Tacet-Discord

python -m venv venv
# Windows:
venv\Scripts\activate
# Linux/macOS:
source venv/bin/activate

pip install -r requirements.txt

3. Start Local Gemma 4 Model (Optional)

If running Ollama locally:

ollama pull gemma4
ollama serve

4. Run the Application

Launch the FastAPI uvicorn application server:

python -m uvicorn main:app --host 127.0.0.1 --port 8000 --reload

Open your browser and navigate to: http://127.0.0.1:8000

---

🌐 API Endpoint Reference

Endpoint	Method	Description
/health	GET	Server health check and agent registration status.
/factory-state	GET	Returns active telemetry baseline, RUL, and memory bank counts.
/junior/ask	POST	Primary Junior Technician Q&A endpoint with semantic matching & LLM grounding.
/records/add	POST	Senior Technician manual record entry with auto-embedding.
/perceive	POST	Visual defect photo perception & heatmap generation.
/correlate	POST	NASA IMS telemetry feature extraction & RUL correlation.
/ask	POST	Classic 5-Agent Pipeline inspection run.

---

🛠️ Project Structure

Tacet-Discord/
├── coordinator/
│   ├── agent.py               # Main Coordinator Agent orchestration
│   └── llm_grounding.py       # Local Gemma 4 & Hosted API grounding engine
├── memory/
│   ├── agent.py               # Hybrid Memory Agent (Vision + Semantic Text)
│   ├── database.py            # SQLite database manager & AI4I 2020 seed loader
│   └── text_matcher.py        # SentenceTransformer all-MiniLM-L6-v2 matcher
├── perception/
│   └── agent.py               # Visual anomaly & heatmap perception agent
├── correlation/
│   └── agent.py               # Telemetry correlation & RUL estimation agent
├── verifier/
│   └── agent.py               # Hallucination Verification Gate agent
├── scripts/
│   └── download_ai4i_data.py  # Dataset loader & preprocessor script
├── static/
│   ├── index.html             # Web Application UI layout
│   ├── app.js                 # Frontend interactions & tab wire-ups
│   └── styles.css             # Industrial dark-mode CSS styling
├── data/                      # Dataset & upload store
├── main.py                    # FastAPI server entry point
├── README.md                  # Project documentation
└── requirements.txt           # Python dependencies

---

📄 License

Distributed under the MIT License. See LICENSE for details.