workflows extract_recipe() in R: Pull the Trained Recipe Out

The workflows extract_recipe() function in R pulls the recipes object back out of a tidymodels workflow, either as the trained recipe carrying learned values like means and medians, or as the original unprepped recipe you attached. It is the standard tool for inspecting, debugging, and auditing what the preprocessor inside a fitted workflow actually learned.

extract_recipe(wf_fit)                              # trained recipe from fitted workflow
extract_recipe(wf_fit, estimated = TRUE)            # explicit, same as default
extract_recipe(wf_fit, estimated = FALSE)           # original unprepped recipe
extract_recipe(wf_fit) |> tidy(number = 1)          # learned values for step 1
extract_recipe(wf_fit) |> bake(new_data = df)       # apply the trained recipe by hand
extract_preprocessor(wf_fit)                        # also works, recipe or formula
extract_mold(wf_fit)$predictors                     # baked predictors after preprocessing

What extract_recipe() does

extract_recipe() returns the recipes object stored inside a workflow. By default it returns the trained recipe, the one prepped during fit() with all learned statistics baked in. Pass estimated = FALSE to get the original, unprepped recipe you handed to add_recipe() instead.

The function exists because a workflow hides its preprocessor inside a slot you should not reach into by hand. wf_fit$pre$mold$blueprint$recipe works, but it is implementation detail that the tidymodels team can change. extract_recipe() is the stable, supported door, and it is part of a wider extract_*() family that does the same job for the model spec, the engine fit, and the baked data.

extract_recipe() syntax and arguments

extract_recipe() takes a workflow and one flag. The signature is short, and the flag is the whole story.

The x argument must be a workflow() object. If the workflow was never fit, extract_recipe() still returns the unprepped recipe regardless of the flag, because no trained version exists. If the workflow uses a formula or variables preprocessor instead of a recipe, extract_recipe() raises an error; reach for extract_preprocessor() in that case.

The estimated argument switches between two snapshots of the same recipe. Both are useful at different points in your modeling cycle.

Pull recipes with extract_recipe(): four examples

Every example below fits a workflow first, then extracts the trained recipe so the focus stays on what extract_recipe() reveals about the prepped pipeline.

Example 1: Inspect learned means and standard deviations

Step_normalize() learns one mean and one standard deviation per column. extract_recipe() plus tidy() shows them.

The numbers come from mtcars and are computed only from rows passed to fit(). Under cross-validation they would be different for each fold, which is exactly the leakage-safe behaviour add_recipe() exists to provide.

Example 2: Bake new data with the trained recipe

A trained recipe can transform new rows directly. No workflow needed once you have the recipe back.

bake() uses the means and standard deviations the recipe learned during fit(). Calling predict(wf_norm_fit, new_data = new_cars) does the same thing internally; pulling the recipe out is only useful when you want the transformed columns without the model on top.

Example 3: Compare estimated = TRUE vs estimated = FALSE

Both snapshots come from the same workflow. The trained one carries learned values; the untrained one carries only the step specifications.

The trained recipe reports the actual learned medians. The untrained recipe reports NA for every value because no data has been seen yet. Pass estimated = FALSE only when you want to reuse the step specs on a different dataset.

Example 4: Reuse a recipe across workflows

Pulling an unprepped recipe lets you wire it into a different workflow without rebuilding the steps from scratch.

Same recipe, different model. The unprepped recipe is the right snapshot here because the random forest will retrain the preprocessing on its own training rows, exactly the same way the linear regression did. Passing estimated = TRUE would freeze the medians learned during the first fit and skip re-estimation.

Common pitfalls

extract_recipe() looks innocent but the estimated flag is easy to misuse. These are the three failure modes you will hit while learning the function.

Try it yourself

Related tidymodels functions

extract_recipe() rarely flies solo. These are the functions you will reach for alongside it.

• add_recipe() attaches the recipe that extract_recipe() will later return.
• recipe() builds the recipes object in the first place.
• extract_preprocessor() works on workflows with any preprocessor type, not only recipes.
• extract_spec_parsnip() pulls the model spec out of the same workflow.
• extract_fit_engine() pulls the underlying engine fit (lm, glm, ranger).
• tidy() summarizes the learned values of any step in the extracted recipe.
• bake() applies the trained recipe to new data by hand.

See the workflows extract reference for the full extract verb family.

FAQ

What is the difference between extract_recipe() and extract_preprocessor()?

extract_recipe() only works when the preprocessor is a recipes object and raises an error otherwise. extract_preprocessor() is the generic version: it returns whatever preprocessor is attached, whether that is a recipe, a formula, or a workflow_variables() selector. Use extract_recipe() when you know the workflow uses a recipe and you want a clear error if that ever changes; use extract_preprocessor() in code that needs to handle multiple preprocessor types.

Does extract_recipe() work on a workflow that was never fit?

It returns the unprepped recipe with a warning. The flag estimated = TRUE cannot deliver a trained recipe when no training happened, so the function falls back to the original recipe. If you want the warning to go away, call fit() first or pass estimated = FALSE explicitly.

Can I save the extracted recipe to disk and reload it later?

Yes. The trained recipe is a regular R object and survives saveRDS() and readRDS() round-trips. Save the trained recipe when you want to apply the same preprocessing to fresh data in a different session without rebuilding the workflow. Watch the version of the recipes package: prepped objects load only against compatible versions.

Why does tidy() on the extracted recipe show NA values?

You passed estimated = FALSE. The untrained recipe stores step specifications without learned parameters, so tidy() reports NA for every value. Drop the flag, or set it to TRUE, and tidy will return the learned means, medians, or dummy levels instead.

How do I see all the columns the recipe produced after baking?

Call extract_mold(wf_fit)$predictors. The mold contains the baked predictor matrix the workflow handed to the model, with one row per training row and one column per output of the recipe. This is useful when a recipe creates many indicator columns and you want to confirm their names before predicting.