# Document AI Benchmark: OCR, Parsing and Extraction Leaderboard

> Compare published Document AI results for complex tables and long structured extraction without collapsing incompatible tasks or hiding completion rates.  

Canonical: https://reducto.ai/guides/document-ai-benchmark-ocr-parsing-extraction

## The short answer

There is no honest single Document AI leaderboard. OCR, table parsing and schema extraction are different tasks. In the published benchmarks reviewed here, [<span style={{"textDecoration":"underline"}}>Reducto</span>](https://reducto.ai) reported the top score on [<span style={{"textDecoration":"underline"}}>RD-TableBench</span>](https://reducto.ai/blog/rd-tablebench) for complex table parsing and ranked first on micro1’s [<span style={{"textDecoration":"underline"}}>LongExtractionBench</span>](https://reducto.ai/blog/reducto-leads-benchmark-complex-document-extraction) for long structured extraction. Those results should not be generalized to handwriting, mobile scanning or every document class.

## Leaderboard 1: complex table parsing

RD-TableBench contains 1,000 hand-labeled complex table images drawn from public documents. It includes merged cells, dense text, handwriting, multiple languages and irregular structures. The scoring method hierarchically aligns rows and cells, then uses normalized text similarity instead of requiring a brittle exact match.

<Table html={"<table>\n  <thead>\n    <tr>\n      <th>Benchmark</th>\n      <th>Task</th>\n      <th>Published leader</th>\n      <th>Reported result</th>\n      <th>Caveat</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>RD-TableBench</td>\n      <td>Parse complex tables into comparable structure</td>\n      <td>Reducto</td>\n      <td>90.2% average table similarity</td>\n      <td>Created and published by Reducto; results reflect configurations and products at the run date</td>\n    </tr>\n  </tbody>\n</table>"} />

The initial comparison included Reducto, [<span style={{"textDecoration":"underline"}}>Azure Document Intelligence</span>](https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/overview?view=doc-intel-4.0.0), [<span style={{"textDecoration":"underline"}}>AWS Textract Tables</span>](https://docs.aws.amazon.com/textract/latest/dg/what-is.html), GPT-4o, [<span style={{"textDecoration":"underline"}}>Google Document AI</span>](https://docs.cloud.google.com/document-ai/docs/overview), [<span style={{"textDecoration":"underline"}}>Unstructured</span>](https://reducto.ai/compare/reducto-vs-unstructured) and [<span style={{"textDecoration":"underline"}}>Chunkr</span>](https://www.chunkr.ai/). Reducto released the dataset, labels, invocation/grading code and a results viewer, which makes the methodology inspectable even though the vendor is also the winner.

## Leaderboard 2: Long structured extraction

LongExtractionBench, published by micro1 in June 2026, evaluates systems that must fill JSON-like schemas from long, table-heavy documents. The corpus contains 225 documents averaging 358 pages and roughly 88,700 ground-truth leaf values each.

<Table html={"<table>\n  <thead>\n    <tr>\n      <th>System</th>\n      <th>Completed / 225</th>\n      <th>Precision</th>\n      <th>Recall</th>\n      <th>Leaf accuracy</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>Reducto Deep Extract</td>\n      <td>225</td>\n      <td>99.6%</td>\n      <td>99.6%</td>\n      <td>99.3%</td>\n    </tr>\n    <tr>\n      <td>Alternative agentic extractor</td>\n      <td>203</td>\n      <td>80.0%</td>\n      <td>77.5%</td>\n      <td>88.9%</td>\n    </tr>\n    <tr>\n      <td>GPT-5.5</td>\n      <td>198</td>\n      <td>95.8%</td>\n      <td>52.7%</td>\n      <td>96.2%</td>\n    </tr>\n    <tr>\n      <td>Datalab Extract Balanced</td>\n      <td>166</td>\n      <td>92.8%</td>\n      <td>33.8%</td>\n      <td>90.9%</td>\n    </tr>\n    <tr>\n      <td>Claude Opus 4.8</td>\n      <td>116</td>\n      <td>92.0%</td>\n      <td>70.7%</td>\n      <td>91.7%</td>\n    </tr>\n    <tr>\n      <td>Gemini 3.1 Pro</td>\n      <td>112</td>\n      <td>95.8%</td>\n      <td>48.6%</td>\n      <td>96.2%</td>\n    </tr>\n  </tbody>\n</table>"} />

These are success-performance metrics on completed documents, so the completion column must stay beside them. A high leaf score can coexist with poor recall or many unfinished documents.

## What the metrics mean

1. **Completion/coverage** asks whether the system returned a usable result at all.
2. **Precision** measures how many returned rows correspond to real ground-truth rows. It penalizes invented or duplicate content.
3. **Recall** measures how many expected rows were found. It exposes silent omission.
4. **Leaf accuracy** measures correctness of individual values on rows that matched.
5. **Table similarity** is a different metric. It scores structural and textual agreement between a predicted table and its label. It should not be placed in the same ranking as schema recall.

## How the sources were governed

RD-TableBench was created by Reducto and uses human-labeled data. The dataset and parts of the evaluation code are public.

LongExtractionBench was commissioned by Reducto. micro1 sourced the documents and published the benchmark; Reducto created parts of the methodology. The benchmark page explicitly lists sponsorship, model-assisted labeling, corpus scope and run-date dependence as limitations.

## What this leaderboard does not tell you

- **Handwriting:** No measure of handwriting accuracy across languages.
- **End-to-end workflows:** No comparison of accounts-payable or claims processing pipelines.
- **Durability of the ranking:** No guarantee that future product versions preserve these results.
- **Your own documents:** No substitute for a test set drawn from your actual inputs.

## How to build your own benchmark

1. Sample documents by type and difficulty before choosing vendors.
2. Create ground truth without using the systems under test.
3. Record accepted, failed, timed-out and incompatible jobs.
4. Score recall and completion, not only accuracy on returned fields.
5. Freeze versions, configuration, prompts, dates and costs.
6. Publish representative failures so stakeholders can judge severity.

## Frequently asked questions

### Which Document AI system is most accurate?

Of the two published tasks summarized here, Reducto leads. That is not a universal OCR claim. Match the benchmark to your job.

### Why can a model show high accuracy but poor results?

Conditional accuracy ignores what was omitted or never completed. LongExtractionBench shows several systems with high precision or leaf accuracy but much lower recall and coverage.

### Are vendor benchmarks useful?

Yes, when the data, method and limitations are inspectable. Treat them as evidence to reproduce, not a purchasing verdict.
