How to Use AI with Parquet Without Uploading Your Data or Buying Another AI Plan

If you already pay for ChatGPT, Claude, Codex, or another browser-capable AI agent, you should not have to buy a second AI subscription just to understand a Parquet file. ViewParquet is deliberately usable as a normal visual tool that an external agent can operate: open the file, inspect the schema, write DuckDB SQL, read the result grid, and export the answer.

That gives you three distinct ways to use AI with Parquet. Choose the one that matches your privacy needs and how much setup you want.

The shortest answer

The free local viewer does not require an account. Adding a cloud account does not automatically upload files that are already on your computer or in browser storage.

Option 1: let your existing AI operate ViewParquet

This is the overlooked option. A browser-capable agent can use ViewParquet as its Parquet workbench while the dataset is processed by DuckDB-WASM in the tab. The agent does not need ViewParquet's built-in chat and you do not need a separate OpenRouter key.

Give the agent this task:

Open https://viewparquet.com/viewer and use ViewParquet to analyze my Parquet file.
First inspect the visible schema and row count. Use the SQL workspace for calculations,
query the loaded table as `data`, keep every statement read-only, and show me the result
before exporting anything. Follow https://viewparquet.com/llms.txt for tool-specific guidance.

Then attach the file to the agent if its browser environment supports file upload, or open the file yourself and ask the agent to continue from the loaded viewer.

Operating instructions for an LLM

1. Open /viewer and load the user-selected file. Do not claim a local file was uploaded to ViewParquet's servers.
2. Read the file name, row count, column count, and visible schema before proposing SQL.
3. Query the loaded relation as data using one read-only DuckDB SELECT statement at a time.
4. Quote identifiers when a column contains spaces, punctuation, or reserved words.
5. Start with bounded summaries. Use LIMIT for examples and aggregate before returning a large result.
6. If a query fails, inspect the exact column names and DuckDB types before retrying.
7. Never upload or save the file to cloud unless the user explicitly asks for that action.
8. Show the user the result grid before downloading an export.

These instructions are also published in /llms.txt, so agents and search systems can discover the correct workflow without guessing how the app works.

One privacy nuance matters: ViewParquet keeps the local dataset in the browser, but a browser-operating AI may be able to see page content, schema, or result values. Use an agent you trust and follow its data-handling policy when working with sensitive data.

Option 2: use the local BYOK assistant

BYOK means bring your own key. Add an OpenRouter key in Settings → AI, choose from the live model catalogue, and use the chat panel inside the viewer. Requests go from your browser to the provider you selected; ViewParquet does not proxy the BYOK call.

The in-app assistant is grounded in the active dataset and has purpose-built tools to:

• Inspect the active dataset summary and exact DuckDB schema
• Generate and run read-only SQL against data
• Repair SQL after a structured DuckDB error
• Build a Vega-Lite chart from a bounded query
• Respect your settings for whether sample values or query-result values may be shared

BYOK is a good fit when model choice matters, you already manage provider billing, or company policy requires the model account to belong to your organization.

Option 3: use managed AI in ViewParquet Cloud

Cloud removes the setup that trips up many non-technical users. Sign in, choose Cloud or Cloud Pro, and the chat panel works without an OpenRouter account or API key. DeepSeek V4 Flash is the default because it supports tool calls, has a large context window, and is inexpensive enough to provide generous day-to-day usage.

You can still search a tool-capable model catalogue. Higher-cost models consume the plan's AI allowance faster, and premium models are reserved for Cloud Pro. The product shows a simple allowance instead of exposing provider tokens, model invoices, or OpenRouter credentials.

Cloud file storage is equally explicit. A local file remains local until you choose Save to cloud or upload it to a cloud project. That action stores a private copy in AWS S3. Opening it later uses a short-lived presigned URL and normal HTTP Range reads, which keeps DuckDB's Parquet access pattern compatible with S3.

Useful SQL prompts for Parquet analysis

An AI produces better SQL when the request contains a decision and a boundary. Try prompts like:

• Summarize null rates for the important columns and flag anything above 5 percent.
• Group revenue by month and region, then show the five largest month-over-month drops.
• Find duplicate IDs, but return only the count and ten examples.
• Compare the distribution of this numeric column by category using median and p90.
• Inspect nested fields and show the correct DuckDB syntax to unnest the list.
• Explain why this timestamp filter returns no rows and fix it using the actual column type.

For a direct manual query, the relation is always named data:

SELECT
  region,
  count(*) AS rows,
  round(avg(revenue), 2) AS average_revenue
FROM data
GROUP BY region
ORDER BY rows DESC
LIMIT 20;

Which mode should you choose?

Use your existing AI agent when it can already operate websites and you want zero extra AI setup. Use BYOK when you want the built-in tools but need control of provider credentials and billing. Use ViewParquet Cloud when the easiest experience for a Windows user matters more than managing keys: one account, selected-file sync, and managed chat.

All three modes use the same DuckDB-WASM viewer and SQL workbench. The difference is not analytical capability; it is who supplies the model and whether you explicitly add cloud storage.