Building an Agentic Analytics Loop with Gemini Function Calling

Most chat-with-your-data demos are a schema pasted into a prompt and a prayer. That's fine until the numbers are real. Brim It triages actual company-card spend, and a wrong total there isn't bad UX. It's a wrong decision someone makes with real money.

So I built around one rule: the model never does arithmetic. A finance manager asks "how much did we spend on travel in Q2, by region?" and the answer always comes from a query. Never from the model's imagination.

a loop, not a prompt

The core isn't a prompt, it's a loop. Gemini gets a small set of read-only query tools and has to call them. It can't touch raw SQL and it can't make up values. It can only ask the data questions through a typed interface.

Five tools, all boring on purpose:

• aggregate — sums/averages over a filtered set
• time_series — values bucketed by day/week/month
• top_merchants — ranked breakdown
• list — raw rows for a filter
• compare — two filtered sets side by side

user: "how much on travel in Q2, by region?"
model -> aggregate({ category: "travel", from: "2026-04-01", to: "2026-06-30", groupBy: "region" })
tool  -> [{ region: "TX", total: 18230.55 }, { region: "CA", total: 11890.12 }, ...]
model -> renders a bar chart + a one-line summary citing those exact totals

Every number on screen traces back to a tool result. The model translates and presents. That's it.

the chart picks itself

The model also returns a viz hint and the client maps it to a component. Bar for grouped aggregates, line for time series, pie for share-of-total, table for raw rows, a single stat for one scalar. Nobody picks a chart type. The shape of the question already implies it.

follow-ups that actually remember

The part that makes it feel like a colleague instead of a search box: follow-ups edit the previous query instead of starting over.

"now just Texas, monthly"

That keeps the travel category and the Q2 window, swaps the groupBy for a monthly time series, and adds region: "TX". The model isn't re-deriving the whole query, it's diffing the last one. The trick is keeping the previous tool call in context, not just the previous answer.

if I built it again

• Read-only, typed tools. "Never invents a number" is a property of the interface, not of prompting harder.
• Give the model a viz channel instead of asking it to format tables in prose.
• Persist tool calls, not just messages. Follow-ups are diffs.

The whole thing is one loop and five small functions. Most of the intelligence is in deciding what the model isn't allowed to do.