data-dictionary-creator

name	data-dictionary-creator
description	Generate a data dictionary for a dataset, combining automatic profiling with the user's description of what the data represents. Use when the user wants documentation of columns — names, types, semantic meaning, units, allowed values, and nullability — for a CSV/Parquet/Excel file.

Data Dictionary Creator

Produce a data dictionary by merging schema inspection with the user's semantic description of the dataset.

Inputs

• Path to a dataset file or folder.
• The user's description of the dataset: what it represents, how it was collected, what each column means (can be partial — infer the rest).
• Optional: output format (markdown default, csv, or json).

Recommended CLI tooling

• duckdb -c "DESCRIBE SELECT * FROM '<file>'" — fast schema + inferred types.
• csvstat — null counts, uniqueness, min/max per column.
• uv run --with pandas python -c '...' — for dtype coercion and sampling.

Procedure

Step 1 — Auto-profile every column

For each column, collect:

• Inferred data type (int, float, string, date, boolean, category)
• Null count and percentage
• Unique count (and full value list if <20 distinct)
• Min / max / mean (numeric) or top-5 modes (categorical)
• Sample values (3 random non-null)

Step 2 — Merge with the user's description

Parse the user's description and map sentences to columns. For each column, fill:

• Name (as in the file)
• Display name / human-readable label
• Description — one-sentence semantic meaning
• Type
• Unit (currency code, SI unit, %, count, etc.)
• Allowed values — enumerated list if categorical with small cardinality
• Nullable — yes/no and what a null means (missing vs. not-applicable)
• Source — where this field originated if the user mentioned it
• PII — none / direct / quasi-identifier (cross-check with pii-flag heuristics)
• Notes — caveats, known issues, derivation formulas

If a column isn't covered by the user's description, mark the Description field as [NEEDS REVIEW] rather than guessing, and list these at the end for user confirmation.

Step 3 — Add dataset-level metadata

At the top of the dictionary:

• Dataset name and path
• Purpose (from user description)
• Row count, column count
• Primary key(s) — infer from uniqueness; ask if ambiguous
• Collection period if derivable from timestamp columns
• Last modified timestamp

Output

Default — write <dataset>-dictionary.md:

# Data Dictionary — <dataset name>

## Overview
...

## Columns

### `column_name`
- **Type**: ...
- **Description**: ...
- **Unit**: ...
- **Nullable**: ...
- **Allowed values**: ...
- **Sample**: ...
- **Notes**: ...

For csv output, flatten to one row per column with standard dictionary columns. For json, emit a structured schema object compatible with JSON Schema / Frictionless Data.

End with a [NEEDS REVIEW] section listing columns the user should clarify.