Run the workflow, read the results, and correct wrong values to make the workflow more and more accurate. The cycle is simple: run the workflow, correct the exceptions during review, and improve the config with Quicktune or the Chat tab. As accuracy improves, the manual work keeps shrinking.
Running a workflow
Select your cases and press Run, and the pipeline executes in order. Your edits are saved automatically as a draft in this browser, and the moment you run, they are committed as a new config version. To create a version without running, press Save on the draft item in the config dropdown at the top. When improving with the Chat tab, you can also commit with the button labeled something like Save as v8 (Improving with the Chat tab).
What one run does
- If you have edited the config, the draft is committed automatically as a new config version (Config #N).
- The selected cases are processed with that committed config.
- Until you run or Save, your edits stay in draft state. See Versions and deployment for how versions behave.
Before you run
Did you select any cases?
Tick the checkboxes of the cases to process in the case list. With nothing selected, the Run button is disabled and reads "Select a document to run your workflow."
Are there any config errors?
If a required setting is empty you'll see "N configuration errors โ click to fix." Clicking takes you to the offending node. For example: an Extract node with no schema, or a Classify node with no document types.
Do you have usage left?
Once you have used all the Studio usage included in your plan, the run doesn't start and you see "You've used all of your Studio usage." Check when usage resets with View usage. To keep running, upgrade your plan or, on Small and above, turn on Allow credit overage (see When usage runs out).
While it runs
- Run status appears in real time in both the case list and the results panel.
- Processing usually takes minutes, depending on document length and pipeline complexity.
- To process everything without selecting cases one by one, choose Run all from the arrow next to the Run button. It processes every case in the list.
Bulk processing belongs on the API. Running from the screen is optimized for testing and review. If you process hundreds of documents a day, see Automating with the API.
"When a run won't start, read the message above the button. The screen tells you whether the problem is file selection, a config error, or usage."
Reading the results
When a run finishes, the config and results panel becomes the results view. Clicking each node in the top node bar to check each stage's results in order is the basic path.
Results by node
| Node | What to check | View formats |
|---|---|---|
| Parse | Whether the document was recognized cleanly as text, and whether tables and headings are intact | Preview / HTML / Text / JSON (Markdown when turned on in Output formats) |
| Classify | Whether documents landed in the intended types, and whether splitting is correct | Grouped list by document type |
| Extract | Whether each field's value is right, and whether any confidence warnings appear | Field list / JSON, with result download ยท for the Mask type, Mask preview / Masked PDF (Mask) |
| Instruct | Whether the response follows the format you asked for, and whether the decision is right | Text response (plus decision value) |
| Validate | Whether each rule passed and the case's traffic-light verdict is as intended (Validate) | Verdict plus per-rule results |
| Merge | Whether the results of the split documents were merged into one (Merge) | Merged result |
| Match | Whether From values and matched values are linked as intended (Match) | From value / Matched value comparison table |
| Export | Saved results are checked in the collection, not on this screen (Collections) | Collection table |
Each node's result can be copied or downloaded as a file from the buttons at the top of the results panel.
If a result looks wrong, start from the earliest node. An empty or nonsensical final result is usually an earlier problem propagating downstream. Walk Parse โ Classify โ Extract โ Instruct and find the first stage that went off.
Switching views
The top of the case panel switches between three views.
- Details: the default working screen, comparing one case's source and result side by side
- Table: the results of many cases laid out in a table at a glance (below)
- Monitor: this workflow's performance metrics (Monitoring)
Table view: reviewing many cases at once
Table view lays the extracted fields of many cases side by side. It is built for bulk review.
- Scan a single field down the column (every invoice date, say). Change the view with Grouping, and narrow it to the cases that match a condition with Add filter.
- Click a cell and the document viewer opens for comparison against the original.
- Edit values directly in the table; your edits become ground truth (below).
- Pick cases with the checkboxes and use Run selected at the top to run only those cases. Rows that have never run get a run button right on the row.
- When downloading as XLSX, you can include per-field confidence columns (grade or score).
"Details for one-by-one checks, Table for bulk review. A click on any cell always opens the original."
Reviewing results and ground truth
Studio's accuracy grows through review. When you fix a wrong value and mark a document as reviewed, that value becomes the ground truth that every later improvement is measured against.
Reading confidence indicators
| Indicator | Meaning | What to do |
|---|---|---|
| No indicator | The model is highly confident | Skim and move on |
| โ Warning triangle | Confidence is below the threshold (this does not mean it's wrong) | Compare against the original |
The confidence threshold is set conservatively, so a warning often sits on a perfectly correct value. Confidence is calculated as
a score between 0 and 1; the API returns it as a low / high grade, and when you download results as XLSX you can choose between
the grade (high/low) and the score (0 to 1). You can turn the warnings on or off with Confidence score under Settings โ My account โ Features.
The three review actions
Edit a value
Double-click a wrong value and fix it (for example, "FedEx" โ "UPS"). The same edit works in table view.
Confirm it's correct
If a warning is showing but the value is right, click the checkmark to confirm "this value is correct".
Mark as reviewed
Once you've checked every field of a document, click Mark as reviewed. Every current value, including the ones you didn't touch, is confirmed as ground truth. You can still edit values afterwards, but the Reviewed state itself can't be undone.
What ground truth does for you
It auto-grades your config changes. Edit a schema and rerun a reviewed document, and the new results are compared against the ground truth automatically. Fields that now match turn green, and fields that are still wrong stay flagged, so you don't re-review everything after every change.
It tracks real accuracy. As ground truth accumulates, you can measure actual accuracy, which is far more trustworthy than confidence scores. The more documents reviewed, the more representative the metric.
Human-in-the-loop operations
For production work that needs human review, we recommend this flow.
- Documents are processed through the API. Processed jobs can be pulled onto the screen from the job list (a job is the processing record of one run) in Monitoring with Import to workflow.
- A reviewer opens each document, verifies and corrects the values, and clicks Mark as reviewed.
- Only the reviewed results are fetched through a separate API call and written to your production database.
Start by reviewing everything, then taper off for the document types whose accuracy has climbed high enough: a gradual path to full automation. That said, for high-stakes documents such as large transactions, keeping review in place regardless of accuracy is the safer call.
"Review is an investment, not a cost. One correction becomes ground truth and then works three times over: auto-grading, accuracy metrics, and Quicktune improvements."
Improving automatically with Quicktune
Quicktune uses the ground truth reviewers have corrected to refine each field's description and build a proposed schema improvement. It doesn't retrain a model; it only changes the extraction descriptions (prompts), so once you apply it, the change goes straight into a new config.
How it works
- Correct wrong values in the results to build up ground truth (Reviewing results and ground truth).
- Click Quicktune in the Extract results panel. It runs in the background and takes up to about 5 minutes. The Classify results panel has the same button, which proposes improved document type descriptions.
- When it finishes, a Quicktune results notification arrives: "We tried improving the schema based on your edits. Check the results."
- Compare the before and after for the fields whose accuracy improved, and apply them with Apply selected keys (Apply selected for Classify). This saves a new config version and reruns all files.
Best practices
- It needs corrected values to start: Quicktune can't run until at least one value has been corrected. The more documents you correct, the better the proposals.
- Always compare before applying: check that a description hasn't been narrowed so much that it breaks existing cases.
- Test before deploying: run the new config version created by applying against your document set before it goes to production.
"Quicktune is a translator from human corrections to schema improvements. Keep reviewing, press the button, and Studio proposes the prompt tuning for you."
Improving with the Chat tab
The Chat tab on the right side of the workflow screen lets you edit the workflow you're looking at through conversation. No digging through settings screens: just ask.
What you can ask for
| Example request | What the assistant does |
|---|---|
Add a 'payment method' field to the schema | Edits the config and applies it as a draft |
Classification keeps getting this wrong. Why? | Inspects the config and results, then proposes a cause and a fix |
Make the documents I just uploaded work too | Looks at the cases uploaded to the workflow and suggests schema and document-type additions |
Run it with the new config | Offers to run, and runs the workflow once you confirm |
Files attached in the Chat tab are for reference in the conversation. Files you attach with the paperclip button in the chat box (attach reference files for chat) are used only in that conversation and are not added to the workflow's run files. Upload the documents the workflow should process as cases on the workflow screen, and talk about them in the Chat tab. To create a new workflow from files attached in chat, use the Home box.
Applying a change
Ask in chat
Start a new conversation in the Chat tab and describe what you want changed. Conversations in this tab are saved against the current workflow.
Check what changed
When the assistant finishes the change, an Updated the workflow configuration card appears, and the changed config is applied to the editor right away as a draft. Expand the card to see the config that was applied.
Save as a new config version
If you're happy with the draft, press the button labeled something like Save as v8 to commit it as a new config version. The config before saving stays in version history, so you can always roll back. When you edit and resend an earlier message, choose Revert & send and the config also goes back to that point.
Chat tab vs. the Home box: use the Home box to create a new workflow, and the Chat tab to improve an existing one. It's the same assistant, so the conversation works the same way.
"If the workflow settings feel complex, hand it to chat. The AI does request โ check โ save for you, and any mistake can be rolled back."