How it works

AskCSV turns plain-English questions about a delimited file into local calculations, tables, and charts. Free analysis needs no account. The distinction is not that the website makes no network requests: it is that your file, questions, SQL, and results are not sent to AskCSV or a cloud AI provider for analysis.

How do I start analyzing a file?

Open the file input, choose a file or paste delimited rows, and check that the columns are recognized correctly. Use a header row with meaningful names, such as region, revenue, and cost.

Before AI analysis, you opt in to an initial model download of approximately 1.8 GB. The model runs through WebLLM and needs a WebGPU-capable browser and compatible hardware. Download time, model initialization, and analysis speed depend on your connection and device; the first answer is not instant.

Later visits may reuse the browser-cached model. Cache storage is not guaranteed: clearing site data, browser eviction, or private-browsing restrictions can mean downloading it again. Local processing is not a promise that the entire website works offline.

Does the AI read every row to calculate an answer?

No. The workflow does not put the entire raw file into the model's prompt. It separates understanding the question from calculating the answer:

1. Understand the columns. The on-device model receives column information and a small number of sample rows to understand labels and value formats. These samples stay on your device.
2. Generate a query. It proposes a read-only SELECT for your question. AskCSV checks the SQL against its allowed-query and column rules before execution.
3. Calculate with DuckDB. DuckDB-Wasm executes the query against the loaded table in the browser, including the rows relevant to its filters and aggregations. Totals come from query execution, not from guessing based on the sample.
4. Explain the result. The app computes summary facts in code and gives the local model a bounded result context to describe. Tables and charts are built from the query result.

Samples, results, and conversation context must fit the local model's limited context window. Ask for focused aggregates rather than expecting it to narrate every raw row.

What does that look like for a sales question?

Suppose a fictional export contains region, gross_sales, and refunds. Ask: "What is total gross_sales by region? Sort from highest to lowest."

A representative calculation would be:

SELECT region, SUM(gross_sales) AS total_gross_sales
FROM sales
GROUP BY region
ORDER BY total_gross_sales DESC;

Here, sales is an illustrative table name, not a command you need to type into AskCSV. The model uses sample rows to understand the fields; DuckDB performs the sum across the loaded table. This is an explanation of the method, not a recorded AI output.

Notice what the question does not specify: subtracting refunds. If you want net sales, ask explicitly for gross_sales minus refunds, summed by region. A valid SQL query can still answer the wrong business question. Check the measure, filters, grouping, and returned numbers before acting on the explanation. For a hand-checkable exercise, use the six-row CSV summary example.

Which file formats and plans are supported?

AskCSV accepts CSV, TSV, and delimited TXT, including pasted delimited text. It does not import .xlsx workbooks directly; export the relevant sheet as CSV first. CSV contains values rather than workbook formatting, multiple sheets, or live formulas.

Free includes 10 successful analyses per month, with no account required. Pro is $5/month and includes unlimited successful analyses, conversation ZIP exports, individual chart PNG exports, deterministic two-file combine, and no ads. Two-file combine uses code to join matching records or stack compatible rows; it is not an arbitrary AI-designed multi-table workflow. See current plan details.

What still uses the network?

Loading the app and downloading the model require network access. Authentication, Stripe billing, analytics, and ads can also make requests, depending on the action, plan, and consent settings. They are separate from the local analysis pipeline.

Local datasets and chats can remain in this browser, so device access, extensions, and local retention still matter. The privacy boundary and synthetic-data network check explain how to evaluate that distinction without using confidential records.

Frequently asked questions

Will it work in any browser?

No. AI analysis requires WebGPU support, compatible hardware, and sufficient device resources. A browser capable of opening the page may still be unable to run the model.

Are answers guaranteed to be correct?

No. SQL validation limits what a query can do; it does not establish that the AI understood your intent. Verify consequential results against a known calculation.

Does local processing guarantee regulatory compliance?

No. It avoids an analysis-content upload, but does not certify GDPR or HIPAA compliance. Your organization must evaluate its device controls, policies, and legal obligations.

Technical references

• DuckDB-Wasm query execution explains how the browser database executes SQL.
• WebLLM getting started documents the WebGPU requirement; advanced usage covers model caching.

Try the workflow with a CSV or pasted sample.