A multi-agent AI system for Brazilian driver's license renewal assistance, built with Google ADK (Agent Development Kit). The system uses a hierarchical agent architecture with an orchestrator routing queries to specialized sub-agents, powered by Vertex AI RAG for document retrieval.

The system follows a hierarchical multi-agent architecture:

orchestrator_agent (routes queries)
├── drivers_license_agent (RAG-powered Q&A)
│   └── Tool: get_drivers_license_context
└── scheduler_agent (clinic search & booking)
    └── Tools: geocode_location, search_nearby_clinics, book_exam

agent/
├── backend/
│   ├── agents/           # Agent definitions (orchestrator + sub-agents)
│   │   ├── orchestrator/ # Root agent that routes questions
│   │   ├── drivers_license/ # RAG-powered license renewal Q&A
│   │   └── scheduler/    # Clinic search and exam booking
│   ├── tools/           # Tool implementations (callable by agents)
│   ├── database/        # Mock database classes
│   ├── rag/             # Vertex RAG pipeline
│   ├── state/           # Shared state key constants
│   ├── types/           # Pydantic models
│   ├── photo/           # Vision API classification
│   └── main.py          # FastAPI server
└── front/               # Flask web UI

Key Features:

• Multi-agent orchestration using Google ADK
• Vertex AI RAG with corpus-based document retrieval
• Mock databases for demonstration (clinics, bookings, photos)
• Photo classification using Google Vision API
• Google Maps integration for clinic geocoding

1. Google Cloud Project with billing enabled
2. Required APIs enabled:
   • Vertex AI API (for Gemini models and RAG)
   • Cloud Vision API (for photo classification)
   • Google Maps API (for geocoding)
3. Service Account with permissions:
   • Vertex AI User
   • Cloud Vision User
   • Storage Admin (for RAG corpus)
4. Python 3.8+

1. Create and activate virtual environment:

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

3. Configure environment variables:

cp .env.example .env

4. Edit .env with your Google Cloud configuration:

GOOGLE_CLOUD_PROJECT_ID=your-project-id
GCS_LOCATION=us-central1
VERTEX_RAG_CORPUS=your-corpus-name
GOOGLE_API_KEY=your-maps-api-key
GCS_BUCKET_NAME=your-bucket-name
VERTEX_AI_INDEX_ID=your-index-id
VERTEX_AI_INDEX_ENDPOINT_ID=your-endpoint-id
EMBEDDING_MODEL=text-embedding-005

5. Authenticate with Google Cloud:

gcloud auth application-default login
gcloud auth application-default set-quota-project <project-id>

# From project root
uvicorn agent.backend.main:app --reload
# Server runs on http://localhost:8000

# Install dependencies (first time only)
cd frontend
npm install

# Start development server
npm run dev
# Web UI available at http://localhost:3000

python run_frontend.py
# Web UI available at http://localhost:5000

python -m agent.backend.agents.orchestrator.agent

Documents must be manually ingested into the RAG corpus before first use:

from agent.backend.rag.document_fetcher import DocumentFetcher
from agent.backend.rag.rag_pipeline import _upload_to_corpus, RAG_CORPUS
from config import DOCUMENT_URLS

fetcher = DocumentFetcher()
docs = fetcher.fetch_all_documents(DOCUMENT_URLS)
for doc in docs:
    _upload_to_corpus(RAG_CORPUS, doc)

Document Sources:

• Lei 15.266/2013 (São Paulo)
• Código de Trânsito Brasileiro (Lei 9.503)
• Resoluções do CONTRAN
• Lei 13.296/2008 (São Paulo)

import asyncio
from agent.backend.agents.orchestrator.agent import call_agent
from agent.backend.types.types import AgentCallRequest

async def test_agent():
    request = AgentCallRequest(
        question="Preciso fazer exame médico para renovar minha carteira?",
        session_id="test-session-1"
    )
    response = await call_agent(request)
    print(response.answer)

asyncio.run(test_agent())

• POST /query - Process user questions
• POST /upload-photo - Upload and classify documents

• Google ADK (Agent Development Kit) - Multi-agent orchestration
• Vertex AI RAG - Document retrieval with corpus-based chunking
• Gemini 2.5 Flash - LLM for all agents
• FastAPI - Backend REST API
• Flask - Frontend web interface
• Google Cloud Vision API - Document photo classification
• Google Maps API - Location geocoding for clinic search

The system uses a custom testing script (no formal test framework):

# Run all tests
python test_agent.py

# Individual test components
python test_agent.py ingestion    # Test document ingestion
python test_agent.py config       # Test configuration
python test_agent.py rag          # Test RAG pipeline
python test_agent.py full         # Full workflow (requires Vertex AI)
python test_agent.py query        # Query only (assumes docs ingested)

• Mock Databases: System uses in-memory mock databases for demonstration
• Portuguese Context: Legal documents and user-facing prompts in Portuguese
• São Paulo Focus: Geographic searches assume São Paulo state/city
• Manual Setup: Documents must be ingested into RAG corpus before first use
• Google Cloud Dependency: Requires active GCP project with proper APIs enabled

Use Google Cloud Application Default Credentials:

gcloud auth application-default login
gcloud auth application-default set-quota-project <project-id>

This project is for educational/demonstration purposes.