Numbers
1.234,56 1,234.56 Compared within tolerance ranges. Trailing-zero, separator and currency-format mismatches don't break the match.
How Invofox measures accuracy in document extraction.
We evaluate, normalize and validate data extraction accuracy across millions of documents — using a transparent, ground-truth benchmarking process built for real-world evals.
Per-field
99.4%
Field-level accuracy
Per-doc
99.2%
Document-level accuracy
Volume
12.4M
Documents evaluated
+18% MoM
Quality
0%
False positives
SLA bound
Powering document extraction for teams at
Ground truth: the starting point for measuring accuracy.
Accurate benchmarking starts with an accurate baseline — what we call the ground truth. It defines the correct data for every field, so we can measure extraction accuracy objectively across your pipeline. When a customer shares their labeled data, we use it as the standard reference.
Document
Ground truth
Comparison
Model output
1{
2 "document_number": "INV-2024-1837",
3 "issued_at": "2024-08-14",
4 "tax_base": 1452.30,
5 "vat_rate": 0.21,
6 "total": 1757.28
7} Ground truth
1{
2 "document_number": "INV-2024-1873",
3 "issued_at": "2024-08-14",
4 "tax_base": 1452.40,
5 "vat_rate": 0.21,
6 "total": 1757.28
7}How Invofox handles complex data in accuracy evaluation.
Document data rarely looks identical, even when it's correct. Our evaluation logic adapts to each data type — so comparisons remain fair and consistent across the board.
Dates
14/08/24 2024-08-14 Normalized to ISO 8601. Time-zone differences and locale formats are reconciled automatically.
Booleans
— false Account for missing or unchecked states. "—" is treated as false unless the schema forces otherwise.
Arrays & tables
[A, B, C] [C, A, B] Evaluated by content, not order — unless the order is business-critical for the use case.
Texts & strings
-
INV-1873INV-1873Exact -
Acme Co.ACME CONormalized -
C/. Mayor 1Calle Mayor 1Levenshtein 0.92
Exact, normalized and similarity-based (Levenshtein) matching depending on the field type.
From field accuracy to full-document reliability.
Most IDP vendors only report field-level accuracy. Invofox goes further by measuring document-level accuracy too — because even a single wrong field can stop an automated workflow. We compute both: per-field for granular analytics, per-document for end-to-end reliability, plus custom validation rules per use case.
End-to-end reliability
Every field must be correct for a document to count. The signal a downstream workflow can actually trust.
Granular analytics
Field-level precision and recall across millions of extracted keys. Perfect for monitoring and dashboards.
Benchmark consistency when your schema evolves.
Adding or removing a field can make old benchmarks impossible to compare. Invofox tracks schema versions and normalizes changes automatically — so your accuracy results stay valid over time. When new keys appear, we flag affected documents so you keep clear visibility into your evolving data model.
-
document_number -
issued_at -
tax_base -
total
-
document_number -
issued_at -
tax_base -
total -
currency
-
document_number -
issue_date -
tax_base -
total -
currency
Accuracy evaluation built on transparency.
We believe accuracy metrics should be verifiable, not subjective. Every eval runs in-house with consistent parameters and transparent rules. Each customer receives both summary metrics and the raw data used to calculate them — no black boxes, no hidden assumptions.
Document number
89.4%
99.3%
Tax base amount
87.9%
98.8%
OrderRef
88.7%
99.1%
Frequently asked questions about accuracy evaluation.
schema.json
1 {
2 ··"question": "What happens if we change our schema mid-test?",
3
4 ··"answer": "Invofox tracks schema versions automatically, aligning field definitions across updates so results remain comparable. You'll always know whether changes come from real performance gains or schema adjustments."
5 }
Still have questions? Talk to us