Ask CSV

What this page does

You have an export. Somebody wants to know which region is ahead, whether last month was up on the month before, and how many rows are missing a value. Answering that means either a pivot table you will rebuild next week, a SQL prompt you have to remember the dialect of, or pasting the file into a chatbot and hoping. This page is the fourth option: type the question, read the number, and see exactly how it was calculated.

Behind the box is a parser with a fixed vocabulary. It peels a sentence apart from the outside in. A date expression comes off first, so "in Q1 2024" is understood as a window rather than mistaken for a column. Then any filter clause, so the word "North" in "where region is North" is never read as a metric. Then the shape of the question: is it a trend, a ranking, a breakdown, a comparison, a data-quality check? Only what is left is matched against your column names, using exact matches, a synonym table that knows revenue and sales and amount are usually the same idea, and edit distance for the typo you made at speed.

What comes out is a DuckDB query. It runs against your file in WebAssembly, in this tab, and the result is drawn with everything you would need to check it: the operation, the matched columns, the literal formula, the result rows, a sample of the underlying rows, and a chart chosen to fit the shape of the answer. If you would rather run the query yourself, the SQL is one click away and it is ordinary SQL.

A worked example

Press Try with sample data and a 10,000-row order file loads: an order id, an order date across two years, a region, a channel, a unit count, an amount and a refunded flag. The left rail fills in immediately. Metrics detected: amount and units. Group by: region and channel. Dates: order_date. Underneath, six starter questions generated from those roles, each one a sentence the parser is guaranteed to answer.

Type total amount by region. The chips read Operation Total amount, Metric amount, Group region. The formula reads SUM(amount) GROUP BY region. The headline is the leading region's total, with the exact figure underneath it, and a sentence saying which region is highest and how many groups there are. A horizontal bar chart appears, then the four result rows, then ten rows from your own file drawn from the leading region so you can see the raw numbers the total was built from.

Now narrow it: total amount by region where channel is online in 2024. Two more chips appear, one for the filter and one for the window, and the SQL grows a WHERE clause with both. Text comparison is case-insensitive on purpose, so typing "online" against data that says "Online" finds the rows rather than returning zero and letting you believe there were none.

Ask amount month over month and you get a line chart of monthly totals with a LAG column showing the change and the percentage change. Ask outliers in amount and you get a count, the fence the count used, and the rows furthest outside it. Ask show missing values and you get one row per column with a count and a percentage, blanks counted alongside nulls because an empty string in a CSV is a missing value whatever the type system says. Ask why did revenue fall and you get a refusal, on purpose.

From answers to a report

Press Pin to brief on each answer worth keeping. Pinned findings collect in the right rail in the order you pinned them, and you can move them up and down until the argument reads properly. Give the brief a title and an analyst note explaining what the reader is looking at, then download it.

The HTML export is one self-contained file. No script tag, no web font, no image, no analytics, and an inline Content-Security-Policy of default-src 'none' that forbids the browser from fetching anything at all, together with noindex and no-referrer directives. Every finding carries its question, its number, its chips, its formula, its result table and, in a collapsed block, the SQL that produced it. It prints properly and it reads properly in dark mode. The Markdown export carries the same content for a wiki or a pull request.

Recipes, for the report you write every week

Most analysis is not a one-off. It is the same eight questions against a new export every Monday. Save the pinned question list as an analysis recipe and you get a small JSON file holding a schema version, a name, a report title, your context note and the questions in order. That is the whole contract: no filename, no rows, no answers, no formulas, no chart points, no column values. You can read the file and confirm it yourself, and the format is documented at recurring report recipes.

Next Monday, drop the new file, press Run on this file, and every question is answered again against the new rows. If a column has been renamed or dropped, the questions that depended on it are skipped and listed by name with the reason. That is deliberate. A recipe that quietly substituted a similar-looking column would produce a report that looks the same and means something different, which is the failure mode worth engineering against.

What it will not do

It will not explain why a number moved. It will not forecast next quarter. It will not recommend an action, name a root cause, or write a narrative around your figures. Those questions are refused by name rather than answered badly. It will also refuse any question whose words do not match a column in your file, and it will tell you which words failed and suggest the nearest column name if one is close. Where two columns fit a phrase equally well, it asks which you meant instead of picking.

For the open-ended question, the full app takes over. There you can bring your own key for any of ten LLM providers, or point it at a local model on your own machine, and the SQL the model writes is validated against your actual schema before it becomes a step you can edit, chain and re-run. The deterministic tier on this page is the floor, not the ceiling.