CSV cleanup & analysis

How to clean a CSV file with CSVian AI

Start with a small, representative sample. Tell CSVian exactly what should change, run the generated tool, and inspect the result before applying it to a larger file. You do not need to write Python yourself.

1. Attach a representative sample

Use the chat attachment control to select your CSV. Tool-building uses up to 10,000 data rows. Include the edge cases you want the tool to handle: blank cells, repeated identifiers, unusual dates and inconsistent spelling.

2. Describe precise rules

For example: “Trim spaces in customer_name. Lowercase email. Keep the latest record for each customer_id using updated_at. Leave blank emails blank.”

Avoid “clean everything” when the correct result depends on your business rules. Explain which columns identify duplicates and how conflicting values should be resolved. For ambiguous dates such as 03/04/2026, specify the intended date order.

3. Run and inspect the result

In Tool, run the transform and open the CSV previews. Compare input and output row counts and inspect changed records. Check that valid rows were not dropped and identifiers retained leading zeros where needed.

Use the audit when you need additional checks. An audit is evidence to review, not a guarantee of correctness. Ask the chat to revise any rule that produces an unexpected result, then rerun the sample.

4. Save a reusable tool

Signed-in conversations appear under Projects. Save a template snapshot when you want to keep a particular version. Open its template page to run it again without rewriting your instructions. Download the finished CSV from the Tool download menu.

Learn a transform from before-and-after examples

Use Learn from an example pair in chat when you already have an input CSV and its intended cleaned version. Supply both, including representative changes and unchanged rows. CSVian analyses the differences to help generate reusable rules.

Test those rules on a different sample. A single example cannot establish every exception: explicitly explain cases such as missing identifiers, conflicting duplicate records or dates that could mean two different things.

Explore charts and the data overview

Open Charts in the workspace or a full template page. It can use the loaded file or the tool output; you can also choose a separate file there.

• Review the overview's data-quality dimensions and reported issues before interpreting a chart.
• Choose Recommend 3 or request a chart using your actual column names, such as “show sales by month”.
• Use a line chart for ordered time values, bars for category comparisons, and scatter plots for two numeric measures. Pie and doughnut charts work best for a few non-negative parts of a whole.
• Download chart images together as a ZIP, or export the overview and charts to PDF. These exports are created locally.

Pro insights adds an AI interpretation of the overview, column names and ten sample values per column. Treat suggestions as hypotheses, especially when the sample is small or unrepresentative. Correlation does not establish causation.

What runs locally, and what is sent?

CSV transforms, the local analysis engines and chart exports run in your browser. Chat and AI-written analysis still use an online AI service. Chat attachments include headers and sample rows; Pro insights sends column samples and the overview. Signed-in projects save conversations and small file samples.

Do not include confidential values in prompts or samples unless you are authorised to share them. Local execution is not the same as an entirely offline or zero-upload app. Keep a copy of your original file and review the output before relying on it.