Installation & Setup

Relevant source files

The following files were used as context for generating this wiki page:

• AI Service/plan.md
• backend/README.md
• backend/test_enhanced_endpoints.py
• frontend/src/pages/admin/InteractionDetail.jsx
• backend/app/utils/prompts.py

This page provides the technical requirements and procedural steps to initialize the AI Anomaly Detection system. The project is designed as a multi-agent watchtower that monitors medical AI outputs, flags safety risks, and provides RAG-grounded explanations. It consists of a FastAPI backend using SQLite for persistence and a React-based frontend for administrative monitoring and human review.

Sources: AI Service/plan.md:1-10, backend/README.md:1-5

Backend Environment Setup

The backend serves as the core orchestration layer, managing the LogAgent, MisbehaviorAnalysisAgent, and the RAGSystem. It requires a Python environment capable of running FastAPI and interfacing with the SQLite database.

Prerequisites & Dependencies

The backend requires administrative configuration and specific environment variables, particularly for the Gemini-based anomaly detection agents.

Component	Description
Python	3.10+ (suggested for FastAPI compatibility)
Database	SQLite (Primary store for logs, scores, and feedback)
LLM Access	API Key for Gemini (for analysis and scoring prompts)
Admin Access	JWT-based Bearer Token for protected endpoints

Sources: AI Service/plan.md:85-92, backend/test_enhanced_endpoints.py:16-24, backend/app/utils/prompts.py:1-5

Administrative Initialization

To access the system's management features, an administrative account must be created. This is done via a dedicated script.

# Create an admin user
python scripts/create_admin.py --email admin@example.com --username admin --password admin123

Sources: backend/README.md:8-11

Frontend Application Setup

The frontend is a React application that provides a "Watchtower UI." This interface allows administrators to view the ML processing pipeline, inspect multi-level anomaly detections, and perform human-in-the-loop review.

UI Architecture

The UI is structured around a multi-stage pipeline view (Levels 1 through 5), ranging from raw ingestion to human verification.

The diagram above shows the logical progression of data visualization within the InteractionDetail component.
Sources: frontend/src/pages/admin/InteractionDetail.jsx:87-290, AI Service/plan.md:195-215

Database & Storage Configuration

The system utilizes a primary SQLite database with specific tables designed for the multi-agent workflow.

Table Name	Description	Key Fields
interaction_logs	Raw user prompts and AI responses	id, prompt, response, model_name
analysis_records	Flags for topics and risk contexts	interaction_id, risk_context_flags
scoring_records	Quantitative risk scores (0.0 - 1.0)	safety_risk, triage_risk, overall_anomaly_score
explanations	RAG-backed reasoning for flags	risk_type, explanation, citations
feedback	Human-labeled audit data	human_label, corrected_response

Sources: AI Service/plan.md:85-95, AI Service/plan.md:13-75

Verification and Testing

Once the environment is configured, the system can be validated using the provided smoke tests and endpoint verification scripts.

Running Smoke Tests

A smoke test is available to verify the integration between the LogAgent and WatchtowerOrchestrator.

# Ensure the server is running before execution
./scripts/phase6_smoke.sh

Sources: backend/README.md:13-16

Enhanced API Testing

For detailed verification of anomaly detection features, use the test_enhanced_endpoints.py script. This requires an active admin token acquired during the setup phase.

# test_enhanced_endpoints.py snippet
BASE_URL = "http://localhost:8000"
API_PREFIX = "/api/admin"
ADMIN_TOKEN = "YOUR_ADMIN_TOKEN_HERE"

headers = {
    "Authorization": f"Bearer {ADMIN_TOKEN}",
    "Content-Type": "application/json"
}

Sources: backend/test_enhanced_endpoints.py:16-24

Summary of Agent Workflow

The installation enables a four-phase pipeline:

1. Data Ingestion: LogAgent captures interactions.
2. Analysis: MisbehaviorAnalysisAgent applies medical safety rules.
3. Scoring: AnomalyScoringAgent calculates safety, triage, and factuality risks.
4. Explainability: RAGSystem generates explanations from medical guidelines.

Sources: AI Service/plan.md:143-155, backend/app/utils/prompts.py:7-140