Skip to main content
Help & Documentation

Import a data dictionary

Load field titles, descriptions, and coded-value meanings from a spreadsheet or survey codebook, from a PDF or other document, or from a documentation page - reviewing every row before anything is applied.

5 min readUpdated October 20, 2026

Most datasets arrive with their meaning somewhere else - a spreadsheet of column definitions, a survey codebook or questionnaire, or a documentation page. Importing a data dictionary moves that knowledge into the dataset itself, so field titles, descriptions, and coded values are readable in the app instead of in a separate file.

Open it

  1. Open the Prepare view of your dataset.
  2. Switch to the "Data Dictionary" tab.
  3. Click "Import Dictionary" in the toolbar.

The drawer is titled "Import Data Dictionary" and offers three routes: drop in a file, paste a documentation URL, or paste codebook text. Clicking the backdrop deliberately does not dismiss it, so a stray click cannot discard work in progress.

From a file

Drag files onto the dropzone, or use Browse files. A dictionary file is a CSV or Excel (.xlsx) spreadsheet. Golden expects four columns - Field Name, Title, Description, and Value Mappings - and is forgiving about headers: Field, Name, or Column all work for the first, and Desc or Definition for the description. A survey-style codebook laid out as Variable / Value / Value Label, with the variable name only on the first row of each block, is read too. A file with no recognizable field-name column is rejected with a message saying so.

You can also drop documents that describe the data - a PDF, Markdown, text, or JSON codebook or questionnaire. Golden's AI reads them and proposes field names, descriptions, coded values, and a dataset summary, landing on the review step described under "From a documentation URL" below. Up to 5 documents totaling 3 MB go in a single import.

Tip: Two links under the dropzone save the setup work: "Download template" gives you a CSV pre-filled with one row per field in your dataset, and "Export current dictionary" gives you what the dataset already knows, so you can edit and re-import it.

Value Mappings is where coded values get their meaning. Write them as pairs separated by commas or semicolons - "0 = No, 1 = Yes" or "M = Male; F = Female" - and Golden reads each pair as one rule.

Review the preview

Nothing is applied until you say so. The preview table lists every row from your file with a check or a cross, and two count pills above it - "N matched" and "N not found" - that double as filters when clicked. A search box filters by field name, title, or description; the All / Matched / Unmatched toggle narrows the same way; and each column header sorts on click and drags on its right edge to resize.

  • Original - the field name as written in your file.
  • Maps to - which dataset field this row will update, editable.
  • Title - the display name to apply.
  • Description - the description to apply.
  • Value Mappings - the parsed rules, shown as pills.

Matching tolerates case, spaces, underscores, and hyphens, so "order_date" finds "Order Date". When it still misses, open the "Maps to" dropdown on that row and pick the right field - the row flips to matched and the Apply count follows. Choosing "Not mapped" leaves the row out. A row you have just remapped stays visible even under a filter it no longer belongs to, until you change the filter again.

The footer button reads "Apply N Updates" and is disabled while nothing matches. Applying reports what it changed - titles, descriptions, and value mappings - and warns separately about any field it could not update.

From a documentation URL

Under "or import from a URL", paste a link to the dataset's documentation page - Kaggle, data.gov, GitHub - and click Extract. Golden reads the page and proposes descriptions and coded values, landing on its own preview step - the same one a dropped document lands on.

  • A confidence pill - high, medium, or low - and how many fields were matched.
  • Dataset Summary and "Coverage & Limitations", both editable text boxes. These become context the AI surfaces can read.
  • A checkbox per proposed field, so you can drop the ones you disagree with.
  • "Replace coded values in data", on by default, when the extraction found any.
Heads up: "Replace coded values in data" rewrites the cells themselves - values like D, S, F become their full names. Turn it off to keep the codes as they are and store the meanings as metadata for the AI only. Read the proposed mappings in the preview before applying, since this is the destructive half.

Paste codebook text

When the codebook lives inside a PDF or Word document, copy its variable list and paste it under "or paste codebook text", then click Extract from text. Golden reads the pasted text the way it reads a document and lands on the same preview. A very long codebook has to be pasted in parts.

Note: On a multi-table model the import is scoped to the table you have selected in Prepare, and the drawer says which one: "Importing into table: <name>". The template, the export, and the "Maps to" list all narrow to that table too, so a row cannot land on a same-named column in a different table.
Tip: An applied dictionary is recorded as a step in your prep history, so it is reviewable and reversible like any other transform.