multi_agent_human_chat

Multi-Agent Insurance Support System

AI agents collaborate in a multi-agent system to provide effective and personalized insurance support.

The code for the example can be found here.

User Interaction

Users interact with the system through a WebSocket-enabled chat interface.

Architecture Overview

Agents Overview

Agents collaborate with clear defined roles, objectives and skills.

FrontendAgent

Role: Serve the frontend and bridge WebSocket communication between web frontend and human communication channel.
Objective: Enable seamless interactions between users and agents through a WebSocket-enabled chat interface.
Skill: Frontend service delivery, Real-time communication, session management, message handling, and agent communication bridge.

TriageAgent

Role: Classify incoming messages and route them to the appropriate agent based on content.
Objective: Ensure that user inquiries are efficiently assigned to the right agent.
Skill: Content classification and routing.

PoliciesAgent

Role: Handle policy-related inquiries using a mock policies_database.
Objective: Provide accurate and detailed information about user policies.
Skill: Policy management expertise.

BillingAgent

Role: Assist customers with billing-related inquiries such as due amounts, billing cycles, and payment statuses.
Objective: Resolve billing-related questions efficiently and provide updates to billing records as needed.
Skill: Billing expertise and data management.

EscalationAgent

Role: Manage support ticket creation and retrieval for escalated issues that other agents cannot resolve.
Objective: Ensure unresolved issues are properly documented and assigned to the correct human support teams.
Skill: Escalation management and ticket tracking.

Communication Flow Example

sequenceDiagram
    participant User
    participant FrontendAgent
    participant HumanChannel
    participant TriageAgent
    participant AgentsChannel
    participant PoliciesAgent

    User->>FrontendAgent: Send greeting (e.g., "hi")
    FrontendAgent->>HumanChannel: Forward user input
    HumanChannel->>TriageAgent: Forward to TriageAgent
    TriageAgent-->>HumanChannel: Respond with initial greeting (e.g., "Hello! How can I assist?")
    HumanChannel-->>FrontendAgent: Send response
    FrontendAgent-->>User: Display response

    User->>FrontendAgent: Ask for help with insurance (e.g., "I want to know what's included in my policy")
    FrontendAgent->>HumanChannel: Forward user input
    HumanChannel->>TriageAgent: Forward to TriageAgent
    TriageAgent->>AgentsChannel: Forward to PoliciesAgent
    AgentsChannel->>PoliciesAgent: Forward to PoliciesAgent
    PoliciesAgent-->>HumanChannel: Ask for clarification (e.g., "Can you provide your policy number?")
    HumanChannel-->>FrontendAgent: Send response
    FrontendAgent-->>User: Display response

    User->>FrontendAgent: Provide policy number (e.g., A12345)
    FrontendAgent->>HumanChannel: Forward user input
    HumanChannel->>TriageAgent: Forward to TriageAgent
    TriageAgent->>AgentsChannel: Forward to PoliciesAgent
    AgentsChannel->>PoliciesAgent: Forward to PoliciesAgent
    PoliciesAgent-->>HumanChannel: Return policy details (e.g., "Your insurance with the policy number A12345 includes...")
    HumanChannel-->>FrontendAgent: Send response
    FrontendAgent-->>User: Display response

Loading

Getting Started

Prerequisites

Ensure you have the following dependencies installed:

• Python 3.12
• Docker and Docker Compose

Ensure you have a valid OpenAI API key set in your environment:

export OPEN_AI_API_KEY="your-api-key"
export GUARDRAILS_TOKEN="your-guardrails-ai-token" # Optional

Setup Instructions

Clone the EggAI repository:

git clone git@github.com:eggai-tech/EggAI.git

Move into the examples/multi_agent_human_chat folder:

cd examples/multi_agent_human_chat

Create and activate a virtual environment and install required dependencies:

make setup

Configure Guardrails (Optional):

guardrails configure --token $GUARDRAILS_TOKEN
guardrails hub install hub://guardrails/toxic_language

Start Redpanda using Docker Compose:

make docker-up

Run the Example

make start-all

Upon running the example and accessing the chat interface at http://localhost:8000, you should see a web-based chat UI.

Agent Evaluation and Optimization

You can execute tests to validate the behavior of different agents using evluation methods like LLM-as-a-Judge.

To run all tests:

make test

Running Specific Agent Tests:

make test-billing      # Runs tests for the Billing Agent
make test-escalation   # Runs tests for the Escalation Agent
make test-frontend     # Runs tests for the Frontend Agent
make test-policies     # Runs tests for the Policies Agent
make test-triage       # Runs tests for the Triage Agent

You can optimize agents using DSPy evaluation and optimization techniques. You can find an optimization setup here.

When running the tests for the Triage Agent, a DSPy performance report will be generated in the agents/triage/dspy_modules/evaluation/reports folder (example report).

Cleaning Up

Stop and remove Docker containers and delete venv:

make docker-down
make clean

Next Steps

• Extend Functionality: Add new agents or tools to handle more complex workflows.
• Connect Real Data: Integrate the system with external databases or APIs.
• Enhance UI: Improve the chat interface with features like authentication and message history.
• Learn More: Explore other examples in the examples folder for advanced patterns.
• Contribute: Share feedback or improvements with the EggAI community.