How to run a Data Exploration analysis

The user types something like:

• "What does the age distribution look like?"
• "Show me the frequency table for brand awareness."
• "Profile household income and education level."

The platform recognises the request as data exploration and picks the tool automatically.

The chat asks the user to confirm which columns to profile. The user can request one column or several at once:

• "Household income"
• "Q5, Q7, and Q12"

The platform maps natural-language column descriptions to actual column names in the dataset. If there's ambiguity, it asks the user to choose from a shortlist.

The chat asks whether the user wants to narrow the population:

• "Only women"
• "Only respondents from 2024"
• "Only people who saw the new packaging"

Filters are applied before any summary is computed, so the statistics and charts reflect only the filtered slice.

The chat shows a confirmation screen:

• Which columns will be profiled
• Any active filters
• The expected sample size

The user confirms, and the profiling runs.

Within seconds, the chat returns — for each requested column:

• Summary statistics appropriate to the variable type:
  • Continuous: count, mean, std, min, quartiles, max
  • Categorical: count, unique values, top 10 frequencies, mode
• A chart:
  • Continuous: histogram with KDE overlay
  • Categorical: bar chart of value frequencies
• Metadata: variable type, sample size, missing-value count

Multiple columns return as a set of side-by-side panels.

The user can continue exploring in the same conversation:

• "Now show me that same variable but only for respondents under 35"
• "What about the satisfaction scores?"
• "Okay, now run a driver analysis on NPS"

The transition from exploration to modelling is seamless — the user doesn't leave the chat or re-upload data.