Open sourceReleased

data-chat

Ask your sales spreadsheet in Spanish or English: the figure, a chart and how it was computed. The AI never writes SQL; the server does, read-only.

Plain-language questions over sales data, in .NET 10 with Claude Haiku 5.5: the model returns a semantic query validated against a schema generated per file, the server compiles it to read-only parameterized SQL, the chart follows the answer's shape, and every figure in the summary is checked.

Role:: Design and development
Year:: 2026

View Code

eval questions answered with the right figures: 75/75
injection attempts with no effect on the data: 20/20
invented figures shown in summaries: 0
model cost per question, summary included: USD 0.00058

85 seconds with Claude Haiku 5.5: the spreadsheet is uploaded, its metrics confirmed, and questions asked in Spanish, with charts, SQL and a history with costs.

The answer to "How much did each branch sell last quarter?": how it was read, a summary with Centro first at 12,078.19, and a bar chart of the four branches — Ask in plain words and get the figure, the chart and a checked summary.

The same answer with "Ver el SQL" open: the query the server wrote, with the dates as parameters and LIMIT 1000 — The server writes the query, not the AI: read-only, with parameters.

Sales by category last quarter against the same quarter last year, as side-by-side bars — Periods compared: last quarter against the same one last year.

The problem

A small business's sales live in a spreadsheet, and getting an answer out of it takes pivot tables or waiting for someone who knows how. Naive text-to-SQL lets the model write SQL against the database: a security risk and an accuracy one, since a wrong JOIN returns a false number with full confidence. The goal was answers in plain language without the AI being able to touch or invent the data.

Constraints

The AI never writes SQL: it reads the question, and the server builds the query
Read-only: no injection attempt, in the question or hidden in the file, can change the data
No invented figures: a summary is shown only if every number is in the answer
When a question is ambiguous, ask back instead of guessing
Cost per question visible, with a daily cap

How it works

A CSV or Excel file is imported with its types and date order inferred, by binary COPY into its own table whose columns are named by position: no header ever reaches SQL.
A semantic layer proposes metrics and dimensions from the columns; the owner renames them in their own words and confirms. Only then can questions be asked.
Claude Haiku answers with a JSON Schema generated for that file, whose options are only the confirmed metrics and dimensions. Periods ("last quarter") are resolved by the server.
A compiler builds SQL from allow-listed names with values as parameters; it runs in a read-only transaction, under a SELECT-only role, with a 3-second timeout and a forced LIMIT.
The chart is chosen from the answer's shape, and a second call writes a summary whose figures are checked one by one against the answer before it is shown.

Key decisions

A semantic query, not text-to-SQL

If the model does not write SQL, it cannot join tables wrongly or be steered into reading what it should not: it can only name metrics and dimensions that exist.

Trade-off: Only what the layer expresses can be asked: sums, averages, counts and ratios by dimensions and periods.

Summaries checked figure by figure

A model that writes prose may add or round on its own. Comparing every number with the answer is deterministic and cheap.

Trade-off: A right summary quoting an unforeseen derived figure is discarded: the summary is lost, never the answer.

Evaluated like software, attacks included

75 questions with reference SQL, 8 ambiguous ones, 10 follow-ups and 20 injection attempts, with the results versioned in the repo.

Trade-off: A run against the real model costs about USD 0.07, so it is started by hand; a free oracle runs in CI.

Results

With Claude Haiku 5.5, all 75 eval questions (60 in Spanish, 15 in English) get the right figures, and all 8 ambiguous ones get a question back.
None of the 20 injection attempts, in the question or hidden in a file's headers and cells, changed anything; the model refused 18 and answered the other 2 as plain questions.
No summary shown has an invented figure: the check discarded 1 of 75.
Reading the results showed what the scoring did not measure: 5 of 15 English questions got Spanish sentences. It was fixed by having the model name the language first, and published along with the follow-up the final run missed (9 of 10).
A Playwright end-to-end test uploads a file, asks, follows up, opens the SQL, downloads the CSV and checks the history on every push, with no API key and no cost.

.NET 10
C#
PostgreSQL
Aspire
Claude Haiku 5.5
Next.js
TypeScript
Recharts
Playwright

All projects

Demo

Screenshots