PARSER GUIDES

Confidence scores

ParsersParser Settings

Understand how confidence scores work, what the numbers mean, and when to use them.

Confidence scores show how sure PARSER is about each extracted value. They help you quickly see which values are safe to trust and which ones might need a human review.

When enabled, confidence scores appear:

• Next to fields in the document view
• In the embed preview when data preview is enabled
• In API responses as numeric values between 0.0 and 1.0

How confidence scores work

When you turn on Calculate confidence scores for extracted fields, PARSER uses a consensus approach:

1. The document is processed three times using the same schema and the same selected extraction model.

2. Each run suggests values for every field and table row.

3. PARSER compares all runs field-by-field, using field descriptions to recognize equivalent formatting or wording.

4. A final value is chosen for each field, and a confidence score is assigned based on how much the runs agree.

At a high level:

• If all runs agree, confidence is 1.0 (100%).
• If two of three runs agree, confidence is about 0.67.
• If the runs disagree a lot, confidence is lower.

If one provider call still fails after retrying, PARSER can calculate confidence from the two successful runs. At least two successful runs are required because one result cannot measure agreement.

Tables and arrays are handled row by row:

• Rows are aligned even if different runs return them in different orders.
• Each cell (e.g. Line Items -> Quantity) gets its own confidence score.
• A row found by only one run is retained so data is not silently lost, but its cells receive a low confidence score.

You don't need to configure any of this--just enable confidence scores in Parser Settings and PARSER handles the rest.

What the scores mean

Confidence scores always fall between 0.0 and 1.0. You'll often see them as percentages in the UI (e.g. 82%).

Interpretation guide

Score

Color in the document view

Suggested action

0.90-1.00

Green

Usually a strong result

0.80-0.89

Blue

Check important business fields

0.70-0.79

Yellow

Review recommended

0.50-0.69

Orange

Review carefully

Below 0.50

Red

Likely needs correction

These colors help you prioritize review; they are not a guarantee that a value is correct. Always check important values such as payment totals, dates, and account numbers.

Enabling confidence scores

To turn on confidence scores for a parser:

1. Open your parser and go to Settings -> Parser Settings.

2. Scroll to Parsing Configuration.

3. Find Calculate confidence scores for extracted fields.

4. Turn the toggle On.

5. Click Save Parser.

Once enabled, new documents processed by this parser will include confidence scores. Existing documents can be re-processed to add confidence scores.

Credit usage

Confidence scores use extra compute because PARSER runs the extraction three times per page and may use smaller reconciliation calls to compare disagreements.

Enabling confidence scores adds +1 credit per page on top of standard processing:

• Standard processing: 1 credit per page
• With confidence scores enabled: 2 credits per page (+1 extra credit per page)

Using the Critical data extraction mode may add more credits. See Credits and Billing for full details.

When your organization uses its own Gemini or OpenAI API key, these model calls are billed directly by that provider. A confidence-enabled extraction normally includes three full extraction runs plus any smaller reconciliation calls needed for disagreements.

Using confidence scores to review documents

Confidence scores are most useful when you're deciding what to review first and what to send straight through.

In the document view

When you open a document in the split view (original document on the left, extracted data on the right):

• Fields and table cells can show confidence visually (for example, as colors or tooltips).
• You can hover or click to see the exact percentage.

Use this to:

• Scan for low-confidence values and fix them directly.
• Focus on documents that have many low-confidence fields.

See Check and Correct Extracted Data for the full review flow.

With Reviews & Approvals

If you use Reviews & Approvals, you can build rules that look at confidence:

• Flag documents when overall confidence is low.
• Flag documents when a specific field has low confidence (e.g. Total Amount confidence < 0.8).

Example rules:

• Confidence < 0.8 - flag documents with confidence below 80%.
• Line Items -> Quantity confidence < 0.7 - flag when line item quantities have confidence below 70%.

Enter review thresholds as decimal values between 0 and 1. For example, use 0.8 for 80%.

See Setting Up Review Stages for how to add rules based on confidence.

Confidence scores in API responses

When you call PARSER's APIs, confidence information is returned alongside your data.

Field-level metadata

In data responses, you'll see a field_metadata object that mirrors your extracted data:

• Each simple field (e.g. email, total) has:
• confidence_score: number between 0.0 and 1.0
• Each array or table has:
• An array of metadata objects, one per item/row
• Each nested field (e.g. education[0].degree, items[1].qty) includes its own confidence_score

This structure lets you:

• Decide which fields to trust automatically.
• Build your own review rules outside PARSER.
• Store confidence scores in your own systems if needed.

Flattened responses

In some exports and integrations, you'll see confidence values:

• As a confidence field directly on each extracted field or table cell.
• As a flattened confidence_score object that follows the same shape as your flattened data.

The numeric meaning is always the same:

• 1.0 = all successful runs agreed, including values recognized as semantically equivalent under the field description
• Lower values = more disagreement across runs

Confidence measures agreement between independent model runs. It is a review signal, not a guarantee that the selected value is correct.

When to use confidence scores

Confidence scores are especially helpful when:

• You process high-value documents (e.g. invoices, contracts, payroll).
• You want to sample and review only risky documents instead of checking everything.
• You're rolling out a new parser and want extra visibility into extraction quality.

They may be less critical when:

• You're processing low-risk documents.
• You already have a strong manual QA process and don't need automated triage.

Tips

• Start with one parser: Enable confidence scores on a single parser first and see how they behave on real documents.
• Use rules, not just eyeballing: Combine confidence scores with Reviews & Approvals rules to automatically flag risky documents.
• Monitor performance: Confidence scores can increase processing time slightly--enable them where they provide the most value.
• Focus on key fields: Pay the most attention to confidence on critical fields like totals, dates, and identifiers.

Next steps

• Parsing Configuration - Other parsing options, including pages and file types
• Check and Correct Extracted Data - How to review and correct extracted values
• Setting Up Review Stages - Flag documents based on confidence and other rules
• Credits and Billing - How credits are calculated