Vertex AI Gemini Agent Chatbot

A deployable AI agent application built with Google Vertex AI Gemini, Firestore, and Cloud Run. It demonstrates persistent conversational memory, controlled tool execution, API design, and practical production-oriented safeguards in one end-to-end project.

Google technology focus

• Vertex AI Gemini for model inference and tool-aware conversations
• Firestore for session-scoped persistent conversation history
• Cloud Run for containerized serverless deployment
• Google Cloud authentication through Application Default Credentials

What this project demonstrates

This repository is designed as a practical reference for developers building agentic applications on Google Cloud. Rather than stopping at a single prompt/response example, it connects the model to memory, tools, an HTTP application layer, and deployable infrastructure.

Core capabilities

• Persistent memory — conversation history is stored per session in Firestore
• Tool calling — the agent can invoke a calculator, a restricted web-fetch tool, and an email integration interface
• REST API — Flask endpoints support chat and session reset flows
• Cloud deployment — the application is structured for deployment to Cloud Run
• Input safeguards — validation, request-size controls, and domain allowlisting reduce unsafe tool execution
• Responsive web UI — a lightweight front end provides an interactive chat experience

Architecture

User
  ↓
Flask Web/API Layer
  ↓
Vertex AI Gemini
  ├── Conversation reasoning
  └── Tool selection
        ├── Calculator
        ├── Restricted Web Fetch
        └── Email integration interface
  ↓
Firestore session memory
  ↓
Cloud Run deployment

Tech stack

Layer	Technology
AI	Google Vertex AI Gemini
Memory	Google Cloud Firestore
Compute	Google Cloud Run
Backend	Python 3.11+, Flask
Front end	HTML, CSS, JavaScript
Auth	Google Cloud Application Default Credentials

Quick start

1. Clone the repository

git clone https://github.com/tejas5038/vertex-gemini-chatbot.git
cd vertex-gemini-chatbot

2. Create a virtual environment

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

On Windows, activate with:

venv\Scripts\activate

3. Configure Google Cloud

Create your local environment file:

cp .env.example .env

Then configure your Google Cloud project and enable the services required by the application, including Vertex AI and Firestore.

Authenticate locally:

gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID

4. Run locally

python app.py

Open http://localhost:8080.

Example API calls

Send a message:

curl -X POST http://localhost:8080/api/chat \
  -H "Content-Type: application/json" \
  -d '{"session_id":"demo-session","user_message":"What is 25 * 47?"}'

Reset a conversation:

curl -X POST http://localhost:8080/api/reset \
  -H "Content-Type: application/json" \
  -d '{"session_id":"demo-session"}'

Deploy to Cloud Run

After configuring your project-specific environment variables, deploy from the repository root:

gcloud run deploy vertex-gemini-chatbot \
  --source . \
  --platform managed \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars GCP_PROJECT_ID=YOUR_PROJECT_ID,GCP_LOCATION=us-central1,FIRESTORE_COLLECTION=chat_sessions

For a real production deployment, review authentication, authorization, rate limiting, logging, secrets management, and access policy before exposing the service publicly.

Security design

The sample includes several safeguards that are useful when building tool-enabled AI systems:

• calculator expressions are constrained rather than passed to unrestricted execution
• web requests are limited to approved domains
• inputs and request sizes are validated
• credentials are expected to remain outside source control
• the email capability is an integration interface and is not enabled with embedded credentials

Why this is useful

This project can be used as a starting point for workshops, demos, and experiments involving:

• Gemini-powered conversational applications
• agent memory patterns
• tool-use orchestration
• serverless AI deployment on Google Cloud
• secure integration of LLMs with external capabilities

Selected upstream open-source contributions

Alongside building Google-technology projects, I contribute fixes and tests to open-source projects in the Google ecosystem.

• Google Magika — PR #1443 — merged upstream
• Google Highway — PR #3361 — merged upstream

Author

Tejas Pravinbhai Patel
Software Development Engineer | AI, distributed systems, and cloud engineering
GitHub: @tejas5038

License

MIT