Skip to main content
Every workflow has a Health tab — the home for measuring how well it performs and driving that up over time. It brings quality metrics, dataset management, and evaluations into one place, so you can measure quality without leaving the workflow.
The short version: the Overview shows how the workflow is doing right now. To measure quality deliberately — and compare workflow versions without the ground shifting underneath you — build a dataset of documents with known-correct answers and run evaluations against it.

​
What’s in Health

Overview

Health at a glance: accuracy and confidence, accuracy over time, and your weakest fields.

Datasets

Manage the workflow’s dataset — add documents, edit ground truth, and tag files into sub-datasets.

Evaluations

Run evaluations, browse the run list, and compare accuracy across workflow versions.

Optimizer

Automatic, iterative improvement of a workflow’s field descriptions against its dataset.
Datasets and evaluations score a workflow’s extraction output, so a workflow needs an Extract step (parse-only and split workflows are excluded). They’re fully decoupled from production — dataset edits and evaluation runs never touch your production data.

​
The Overview

The Overview gives you a high-level read on workflow health:
  • Average confidence and average accuracy
  • Accuracy and confidence over time
  • The weakest fields — where errors concentrate
New to these two numbers? See Analytics & Quality for what each one means and how to use it.

​
The quality loop

Health is built around a simple, repeatable loop:
1

Build a dataset

Add representative documents and confirm their correct values as ground truth. See Building a dataset.
2

Evaluate

Run a workflow version over the dataset and read its accuracy. See Running evaluations.
3

Refine

Improve fields, instructions, or the pipeline based on where it’s failing — or let the Optimizer rewrite field descriptions for you.
4

Re-evaluate and compare

Run again on the same dataset. Because the dataset is fixed and each evaluation is immutable, the difference between two runs is real signal, not noise.

​
What’s next?

Building a dataset

Add documents and author ground truth.

Running evaluations

Score a version and compare across versions.