2. From Question to Answer: The Six-Step Pipeline

When you type a question like "total revenue by region for Q2," OUPI doesn't hand your data to a language model and hope for the best. Instead, it follows a precise six-step pipeline:

1. Code generation — The model reads your dataset's structure (columns, types, a sample of rows) and writes executable code that answers your question.
2. Security check — That code is inspected before it runs: no network access, no file-system reach beyond your dataset, and the question itself is screened for manipulation attempts.
3. Sandboxed execution — The validated code runs in an isolated environment on the full dataset — every row, not a sample.
4. Plain-language summary — The model writes a human-readable answer from the computed result.
5. Number verification — Every figure in the summary is compared against the actual computed data. Any mismatch is flagged with a warning.
6. Chart suggestion & follow-ups — If the result shape suits a visualization, a chart is proposed. Related follow-up questions are also suggested to keep your analysis moving.

Arithmetic is always done by code, never approximated by a model. That's the core reliability guarantee.

Why code, not a model doing math?

A language model reading 10,000 rows will approximate — it might round, skip, or hallucinate totals. Code counts exactly. After the code runs, the summary is checked against the computed figures, so a wrong number cannot slip in unnoticed. What can vary is how your question was interpreted, which is why OUPI shows you the generated code and the result table alongside the summary. If the interpretation is off, you can adjust your wording and re-ask.

Security at every gate

Before any generated code touches your data, it passes through multiple guards. Suspicious patterns are blocked and logged. The question itself is screened for prompt-manipulation attempts. Execution then happens inside a sandbox — isolated from the network, the file system, and the rest of the platform. Your data never leaves the platform during this processing. This means you can query sensitive business data (sales figures, margins, HR metrics) without worrying about leakage.

What you keep after each query

Every query stores four artifacts: the generated code, the result table, the plain-language summary, and the suggested chart. Save a query to re-run it later on refreshed data — when your dataset is updated, the same logic runs on the new rows. Saved queries are shareable with your team, so a well-crafted question becomes a reusable analytical asset, not a one-off answer.

Know the limits

The engine answers questions the data can answer. It will not invent missing columns or fill gaps with assumptions. For very large datasets, the model samples the structure to understand it, but the computation still runs on all rows. If your analysis requires complex, multi-step modeling (forecasting, scenario building), consider launching a mission where @oupi can iterate over several steps rather than answering in a single query.