How to ask a local AI model about your CSV or Excel data

Pasting a spreadsheet into ChatGPT gets you an answer fast. It also hands your data to someone else. A local AI model is an open-weight LLM running on your own GPU. It gives you most of the same convenience, and your data stays put. Here’s how it works in the browser and how to get good answers out of it.

What “local AI model” means here

NoEgress runs open-weight models like Qwen 2.5, Llama 3.2, and Gemma 2 through WebLLM, which uses WebGPU to run them on your graphics card. The model downloads once into your browser cache. That’s about 1 GB for the smallest and about 5 GB for the largest. After that it works offline.

No API key, no provider, no bill. Everything runs in a web worker inside your tab.

The model writes SQL, DuckDB does the math

Small language models are bad at arithmetic, so the assistant doesn’t do any. It gets the table schema, some column stats, and a few sample rows, then writes a DuckDB query. DuckDB runs that query over the whole file and gives you an exact answer.

That’s why you can trust the number. It came from SQL you can read and edit, not from a model guessing at totals over rows it never saw.

-- "Which three regions grew revenue the most this year?"
SELECT
  region,
  SUM(revenue) FILTER (WHERE year = 2026)
    - SUM(revenue) FILTER (WHERE year = 2025) AS growth
FROM orders
GROUP BY region
ORDER BY growth DESC
LIMIT 3;

Choosing a model

Start with Qwen 2.5 3B, the default. It loads quickly and writes solid SQL. If your GPU has the memory, Qwen 2.5 7B or Llama 3.1 8B handle harder, multi-step questions better. On an older laptop, Llama 3.2 1B still manages simple lookups.

Getting better answers

Use column names as they appear in the file, and say what you want measured. "Sum of revenue by region" beats "how are we doing?" If a query fails, ask the assistant to fix it. It gets the error message along with the SQL.

Check the Data Profile first. If a number column came in as text, no model will add it up correctly until you cast it.

Local AI vs ChatGPT for spreadsheets

A hosted model is bigger and writes better prose. A local one is private, free, and works offline. It’s also allowed on data your company won’t let you paste into a third-party service. For turning a question into a query over a table you already have, a 3B to 8B model is usually plenty.