How to chat with Google BigQuery in under 5 minutes

BigQuery is one of the easiest warehouses to query — once you know SQL. For everyone else, the SQL editor is a wall. This tutorial walks you through pointing Data Talks at a BigQuery dataset and asking your first natural-language question. End to end, it takes less than 5 minutes.

Step 1 — run Data Talks locally

Clone the repo and start the stack with Docker Compose. You'll get a UI on http://localhost:5173 within about a minute.

git clone https://github.com/Empreiteiro/data-talks.git
cd data-talks
docker compose up -d

Step 2 — create a BigQuery service account

In Google Cloud Console, create a service account with two roles: BigQuery Data Viewer (read tables) and BigQuery Job User (run queries). Generate a JSON key for it. This is the credential Data Talks will use.

Tip: scope the service account to a single dataset if you want strict isolation. Use IAM conditions on the dataset, not the project.

Step 3 — add BigQuery as a source

In the Data Talks UI, click 'Add source' → 'BigQuery'. Paste the service account JSON, pick the project and dataset. Data Talks introspects the schema in seconds.

Step 4 — ask your first question

Open a new conversation. Pick the BigQuery source. Ask something simple like:

Show me total revenue by month for the last 12 months.

Within a few seconds, the agent generates Standard SQL, runs it, and returns a line chart. Click 'Show SQL' to inspect the query — it will look like a sensible monthly aggregate over your fact table.

Step 5 — pin the chart and set an alert

Pin the chart to a dashboard. Then click 'Set alert' on the conversation: pick weekly, configure a webhook or Telegram destination, and the same question now runs every week with results delivered to your team.

Where to go next

• Add a second BigQuery dataset and ask a cross-source question.
• Connect your Postgres replica too — Data Talks can join across sources in a single conversation.
• Try Studio reports for executive-style summaries with AI commentary.
• Read the integration page for BigQuery — it covers cost control, partitioned tables, and JOIN inference.

That's the loop: connect → ask → pin → alert. Five minutes from clone to a recurring metric in your inbox.