Extraction accuracy is the share of extracted field values that match a verified ground truth. In commercial real estate document extraction, it measures how often the values pulled from a lease, offering memorandum, or rent roll are correct. It is always reported against a specific reference set and at a specific level, most often per field.
How Is Extraction Accuracy Measured?
Extraction accuracy is measured by comparing each extracted value to a verified ground truth and counting how many match. The result depends entirely on the level of comparison chosen. The same extraction can score far differently at the character level, the field level, or the document level, so a stated accuracy figure is meaningless without its level.
Three levels are standard:
Character-level, tracked as Character Error Rate (CER), the share of individual characters wrong, computed by Levenshtein edit distance. Clean printed text benchmarks below 1% CER, or above 99% character accuracy (LlamaIndex).
Field-level accuracy, the share of complete field values correct. For business use 95%+ is acceptable and 99%+ is excellent (LlamaIndex).
Document-level, the share of documents with every field correct, the strictest and lowest of the three.
Level | Metric | Representative benchmark |
Character | Character Error Rate | Below 1% CER on clean printed text |
Word | Word Error Rate | Below 2% WER on standard documents |
Field | Field-level accuracy | 95%+ acceptable, 99%+ excellent |
Document | All-fields-correct rate | Strictest; lowest of the four |
Why Extraction Accuracy Matters
Extraction accuracy matters because the threshold for removing human review is set in the last fraction of a percent, and each level tells a different story. Character accuracy above 99% can still leave field accuracy well below it, because a single wrong character breaks the whole field. The metric that governs a workflow is field-level, since that is the unit an operator underwrites on.
The practical stakes sit at the top of the scale. For financial and identity fields, the benchmark to enable straight-through processing, where documents move through the workflow with no human review, is 99.9% field-level accuracy (IDP Software). Below that, a human-in-the-loop step is required to catch the residual errors. The jump from 99% to 99.9% is not cosmetic: at 99% field accuracy, 1 field in 100 is wrong, which across a 60-field lease means most leases carry at least one error. Extraction accuracy is the number that decides whether a reviewer is optional or mandatory.
Example
Extraction accuracy is clearest computed at two levels on the same run. A model extracts 40 fields from an offering memorandum. The base rent field, "$32,500.00", is read as "$32,600.00": every character but one is correct, yet the field value is wrong. Across the full document, 38 of 40 fields match ground truth exactly.
Level | Correct | Total | Accuracy |
Character (base rent field, 10 chars) | 9 | 10 | 90.0% |
Field (whole document) | 38 | 40 | 95.0% |
The calculation walkthrough:
Character accuracy on the base rent field = 9 / 10 = 0.900, or 90.0%, since one digit of ten is wrong.
Field-level accuracy = correct fields / total fields = 38 / 40 = 0.950, or 95.0%.
The base rent field is 90% correct at the character level but counts as a full failure at the field level, because "$32,600.00" is not "$32,500.00". This is why field-level accuracy is the honest metric for underwriting: a value that is one digit off is not 90% useful, it is wrong. The 95.0% field accuracy here, two errors in forty, would route to human review before the numbers reached a model.
Variations and Edge Cases
Extraction accuracy changes with the level measured and the matching rule applied. The variants below shift what a single accuracy figure certifies.
Variant | Behavior |
Character-level (CER) | Forgiving; a near-miss still scores high |
Field-level | The value is correct only if every character matches; the operator metric |
Exact vs fuzzy match | Fuzzy matching inflates accuracy by accepting near-values |
Weighted accuracy | High-stakes fields, like rent, counted more heavily than minor ones |
Straight-through-processing threshold | 99.9% field accuracy required to skip human review |
Extraction Accuracy vs Precision and Recall
Extraction accuracy is often confused with precision and recall, but it answers a different question. Extraction accuracy is the overall share of values that are correct. Precision and recall split that outcome into two failure modes: precision measures how many extracted values were correct, and recall measures how many present values were captured.
Put simply, accuracy is the headline and precision and recall are the breakdown. A model can post 95% accuracy while hiding whether its errors are wrong values or missed values. If it silently skips fields, recall falls while accuracy on what it returns stays high. Accuracy alone will not surface that, which is why extraction is reported on accuracy plus precision and recall together.
Frequently Asked Questions
What is a good extraction accuracy for document processing?For field-level accuracy, 95%+ is generally acceptable and 99%+ is considered excellent (LlamaIndex). To enable straight-through processing with no human review on financial or identity fields, the benchmark is 99.9%. The right target depends on the cost of a single wrong field.
Why is field-level accuracy lower than character accuracy?Because a field is correct only if every character in it is correct. A model can read 99% of characters right and still get a field wrong when the one bad character falls inside it. Character accuracy above 99% routinely maps to lower field-level accuracy.
How is extraction accuracy calculated?At the field level, extraction accuracy is the number of correctly extracted fields divided by the total number of fields, measured against a verified ground truth. At the character level it is one minus the Character Error Rate, the share of characters wrong by edit distance.
Related Terms
Ground Truth
Precision and Recall
Optical Character Recognition
Intelligent Document Processing
Confidence Score