Multi-Agent Insurance Support System

A production-ready reference architecture for building enterprise AI agent systems. Features 7 collaborating agents, 8 classifier strategies, RAG-powered document search, and full observability.

Quick Start

# Clone and start (one command!)
git clone git@github.com:eggai-tech/eggai-demo.git
cd eggai-demo
make start

Open http://localhost:8000 and start chatting.

Note: Runs completely locally with LM Studio - no cloud services or API keys required!

What This Demo Shows

Feature	Description
7 Collaborating Agents	Triage, Billing, Claims, Policies, Escalation, Audit, Frontend
9 Classifier Strategies	Compare LLM, fine-tuned, and neural network approaches (v0-v8)
RAG Document Search	Vespa-powered hybrid search (70% semantic + 30% keyword)
Production Patterns	Health checks, observability, message-driven architecture
Full Observability	Grafana dashboards, distributed tracing, metrics

Architecture at a Glance

User (WebSocket) → Frontend Agent → Triage Agent → Specialized Agent → Response
                                         │
                                    Classifier
                                    (v0-v8: pick your strategy)

Message Flow:

1. User sends message via WebSocket to Frontend
2. Triage classifies intent using configurable classifier (v0-v8)
3. Routes to specialized agent: Billing, Claims, Policies, or Escalation
4. Audit monitors all interactions for compliance
5. Response streams back to user

Commands

Command	Description
make start	Start everything (infrastructure + agents)
make stop	Stop agents
make health	Check service health
make test	Run tests
make help	Show all commands

Explore the Codebase

Entry Points

Want to...	Start here
Understand the system	docs/system-architecture.md
See how agents work	agents/triage/agent.py
Compare classifiers	agents/triage/classifiers/
Add a new agent	docs/building-agents-eggai.md
Configure RAG search	docs/agentic-rag.md

Code Organization

├── agents/                    # AI agents (self-contained modules)
│   ├── frontend/             # WebSocket gateway + chat UI
│   ├── triage/               # Classification and routing
│   │   ├── agent.py          # Message handler
│   │   ├── classifiers/      # 9 classifier strategies (v0-v8)
│   │   └── dspy_modules/     # DSPy-based classifiers
│   ├── billing/              # Payment inquiries
│   ├── claims/               # Claim processing
│   ├── policies/             # RAG-powered policy search
│   ├── escalation/           # Complex issue handling
│   └── audit/                # Compliance monitoring
├── libraries/                 # Shared utilities
│   ├── communication/        # Kafka channels
│   ├── observability/        # Logging, tracing, metrics
│   └── ml/                   # DSPy, MLflow integration
├── config/                    # Configuration
│   └── defaults.env          # Sensible defaults (works out of box)
├── scripts/                   # Startup utilities
│   ├── start.py              # One-command startup
│   ├── stop.py               # Graceful shutdown
│   └── health_check.py       # Service health checks
└── docs/                      # Documentation

Classifier Comparison

The triage agent supports 9 classification strategies. Select via TRIAGE_CLASSIFIER_VERSION:

Version	Type	Latency	API Call	Training
v0	Minimal prompt	~500ms	Yes	No
v1	Enhanced prompt	~600ms	Yes	No
v2	COPRO optimized	~500ms	Yes	One-time
v3	Few-shot MLflow	~50ms	No	Yes
v4	Zero-shot COPRO	~400ms	Yes	One-time
v5	Attention network	~20ms	No	Yes
v6	OpenAI fine-tuned	~300ms	Yes	Yes
v7	Gemma fine-tuned	~100ms	No	Yes
v8	RoBERTa LoRA	~50ms	No	Yes

Default: v4 (configured in config/defaults.env — best balance of accuracy and simplicity)

# Using the unified classifier interface
from agents.triage.classifiers import get_classifier, list_classifiers

classifier = get_classifier("v4")
result = classifier.classify("User: What's my bill?")
print(result.target_agent)  # BillingAgent

# Compare all classifiers
for info in list_classifiers():
    print(f"{info.version}: {info.name}")

Infrastructure Services

Service	URL	Description
Chat UI	http://localhost:8000	Main application
Redpanda Console	http://localhost:8082	Message queue UI
Vespa	http://localhost:8080	Vector search
Temporal UI	http://localhost:8081	Workflow monitoring
MLflow	http://localhost:5001	Experiment tracking
Grafana	http://localhost:3000	Dashboards
Prometheus	http://localhost:9090	Metrics

Configuration

Configuration uses a 3-layer approach:

1. config/defaults.env - Sensible defaults (committed, works out of box)
2. .env - Local overrides (gitignored)
3. Environment variables - Runtime overrides

Key settings:

# Classifier selection
TRIAGE_CLASSIFIER_VERSION=v4

# LLM provider (default: local LM Studio)
TRIAGE_LANGUAGE_MODEL=lm_studio/gemma-3-12b-it-qat

# Or use OpenAI
# TRIAGE_LANGUAGE_MODEL=openai/gpt-4o-mini
# OPENAI_API_KEY=sk-...

Requirements

• Python 3.11+
• Docker and Docker Compose
• uv (recommended) or pip
• LM Studio (for local models) or OpenAI API key

Development

# Run tests
make test

# Run with coverage
make test-coverage

# Lint code
make lint

# Auto-fix lint issues
make lint-fix

# Full reset (removes all data)
make full-reset

Documentation

• System Architecture
• Agent Capabilities
• Multi-Agent Communication
• Building Agents
• RAG with Vespa
• Classifier Guide
• Deployment Guide

License

MIT