How AskCSV works

From question to answer

1. 1

   Load the data in your tab

   Drop a CSV or pick a sample. DuckDB-WASM, a SQL engine compiled to WebAssembly and loaded from the jsDelivr CDN, reads the file into one table inside this browser tab. It then profiles the table with SUMMARIZE and takes 5 random sample rows.

2. 2

   Plan a query

   Your question and the table profile go to the AskCSV server, which asks an OpenAI model for a typed plan: a one line intent, one DuckDB query and a chart spec.

3. 3

   Check the SQL is read-only

   The server rejects anything that is not a single SELECT or WITH statement, any write or DDL keyword, PRAGMA, SET, ATTACH, INSTALL, and functions that reach outside the table such as read_csv, read_parquet or glob. The browser runs the same check again right before executing, including SQL you edit by hand.

4. 4

   Run it, and repair if needed

   The query runs in DuckDB in your tab. If it errors, is blocked or returns no rows, the SQL and the error go back to the model for a fix, up to 3 attempts in total. Every attempt is listed under "How I got this".

5. 5

   Draw the chart

   The chart spec is checked against the columns the query actually returned. Unknown or non-numeric columns are dropped, scatter plots need a numeric x axis, and a chart with too many points falls back to a table.

6. 6

   Write and check the answer

   Up to 200 result rows go to the model, which writes one to three sentences. Every number in the answer is matched against the result cells, allowing for rounding such as 38.2k. If a number is not found, the server asks once more. The browser repeats the check and shows a warning naming any number it could not find.

What runs where

What the model sees

The CSV file itself never leaves your device. These parts of it do, through the AskCSV server to OpenAI:

• To write SQL: your question, the table name and row count, column names and types, per-column stats (min, max, average, distinct count, share of nulls, and for text columns with 40 or fewer distinct values the 12 most common values with counts), and 5 sample rows with text cut to 120 characters.
• For follow-ups and repairs: the last 4 questions with their intent and SQL, and the SQL and error of failed attempts.
• To write the answer: your question, the intent, the SQL, the result column names and up to 200 result rows.

The AskCSV server does not save any of this. It only records when each IP address asked, to enforce the hourly limit. What OpenAI keeps is set by its API data policy, not by this app. Your conversations, including up to 200 result rows per answer, stay in this browser's local storage until you delete them. Uploaded files are not kept, so you drop the file again to continue an old conversation.

An example

On the SaaS subscriptions sample, the question "What is active MRR by plan?" leads to a query like this one. The model writes its own SQL each time, so yours may differ.

SELECT plan, ROUND(SUM(mrr_usd), 2) AS active_mrr
FROM subscriptions
WHERE status = 'active'
GROUP BY plan
ORDER BY active_mrr DESC

The model is told to chart a few categories like these as bars.

Sample datasets

• SaaS subscriptions

  Plans, seats, MRR and churn for a B2B tool, 2024 to 2025.

  3,200 rows

  • "What is active MRR by plan?"
  • "Show monthly signups in 2025 split by billing cycle"

• City bike trips

  A year of trips with stations, rider type, bike type and weather.

  4,200 rows

  • "How many trips per month, members vs casual riders?"
  • "Which start stations are busiest?"

• Coffee shop sales

  Order lines from three stores, January to September 2025.

  4,800 rows

  • "Revenue by store per month"
  • "What are the top 5 products by revenue?"

Limits

• Files: .csv, .tsv or .txt, up to 25 MB, with a header row.
• One table at a time. Loading a new file replaces the previous table, so joins across files are not supported.
• The model sees a profile of the first 80 columns.
• Questions can be up to 500 characters. Follow-up questions carry the last 4 answered turns as context.
• Up to 3 query attempts per question.
• The browser reads up to 5,000 rows of a query result and keeps 2,000 for the chart and table.
• Charts hold up to 60 bars, 400 line points or 2,000 scatter points before falling back to a table.
• The free demo has an hourly question budget per IP address (10 by default in the code). The number left is shown next to the question box.