Chatty Data Generation

A conversational AI app with two functions:

1. Synthetic data generation - parse a SQL DDL schema and generate valid synthetic data that respects all constraints (especially foreign keys).
2. Talk to your data - query the generated data in natural language, with results rendered as text, tables, and plots.

Show "Talk to your data" tab screenshot

Tech stack

• LLM: Gemini 3.5 Flash (2.0+ supported) - function calling, structured/JSON output.
• SDK: Google GenAI SDK (Vertex AI auth via a GCP project).
• UI: Streamlit.
• DB: PostgreSQL.
• Container: Docker.
• Observability: Langfuse.

Project layout

src/        application source code
examples/   sample SQL schemas to try
assets/     screenshots used in this README
tests/      test suite

Quick start

Everything (app + PostgreSQL) runs in Docker. Authentication uses Vertex AI via Application Default Credentials - no plain API keys. Prerequisites: Docker, the gcloud CLI, and access to a GCP project with Vertex AI enabled.

1. Clone and enter the repo.

   git clone <repository-url>
   cd Chatty-data-generation

2. Authenticate to Google Cloud. This writes the Application Default Credentials that the container mounts.

   gcloud auth application-default login

3. Configure your environment.

   cp .env.example .env

   The Vertex AI settings (project, location) live in docker-compose.yml. It defaults to project gd-gcp-gridu-genai; to use your own, edit GOOGLE_CLOUD_PROJECT there. .env only carries optional Langfuse keys.

4. Run it.

   docker compose up --build

   Once it's up, open http://localhost:8501. Stop with Ctrl+C.

Tip: sample schemas live in examples/ (library_mgm.ddl, restaurants.ddl, company_employee.ddl). Upload one in the Data Generation tab to try it out.

Note: data refinement is per-table - each feedback box regenerates only the selected table. A single global edit across all tables at once (e.g. "replace X with Y everywhere") is not supported; apply such changes one table at a time.

Configuration

Full reference for the keys in .env. The Docker quick start authenticates via Vertex AI (settings baked into docker-compose.yml); these keys matter mainly for the non-Docker dev path.

Key	Purpose
GOOGLE_GENAI_USE_VERTEXAI	true (default) for Google Cloud (Vertex AI) auth; false only for the dev-only API-key path
GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION	GCP project + region for Vertex AI auth
GEMINI_API_KEY	Gemini API key, used only when GOOGLE_GENAI_USE_VERTEXAI=false (local dev only, not spec-conformant)
GEMINI_MODEL	Model id (default gemini-3.5-flash)
LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY / LANGFUSE_HOST	Optional observability (leave blank to disable)

Auth: the project standard is Vertex AI (no plain API keys). Run gcloud auth application-default login once so the credentials are available. The plain API-key route (GOOGLE_GENAI_USE_VERTEXAI=false) is a local-dev convenience only and does not work with the Docker setup, which forces Vertex AI.

Running for development (without Docker)

For working on the code directly. Requires uv and a local PostgreSQL database.

uv sync --extra dev          # create .venv and install dependencies
cp .env.example .env         # then fill in your credentials
uv run streamlit run src/app.py

Run tests with uv run pytest.