---
title: "Extend: Document Processing Infrastructure for AI Agents"
description: "Parse, extract, and split your hardest documents with unmatched accuracy. Read any layout with specialized vision models, and ship reliable pipelines in minutes, not months."
canonical: https://www.extend.ai/
---

# Extend: Document Processing Infrastructure for AI Agents

**New** — Introducing Light Parse: accurate parsing at a fraction of the cost [Read blog](https://www.extend.ai/resources/introducing-light-parse)

## Production-ready document processing

Parse, extract, and split your hardest documents with unmatched accuracy. Ship reliable document agents in minutes, not months.

[Try for free](https://dashboard.extend.ai/) · [Book demo](https://www.extend.ai/get-started)

### Install

| SDK | Command | Docs |
| --- | --- | --- |
| Python | `pip install extend-ai` | [View docs](https://docs.extend.ai/developers/sdks) |
| Typescript | `npm install extend-ai` | [View docs](https://docs.extend.ai/developers/sdks) |
| CLI | `curl -fsSL https://extend.ai/install.sh | sh` | [View docs](https://docs.extend.ai/cli) |

## Trusted by leading AI teams

CH Robinson, Flatiron, Brex, Checkr, Square, Amgen, Mercury, Opendoor, FactSet, Ironclad, First American, Comulate, Axle, SamaCare, Pallet, Savvy, Nuvocargo, Collective, Valon, Upstart

## The highest performance document APIs for your agents and pipelines. Processing millions of pages every day.

APIs:

### 01 / parse

Convert unstructured documents into context for agents.

- **Layout detection** — Advanced layout model detects tables, checkboxes, images, handwriting, and signatures on every page.
- **Specialized vision models** — A hybrid computer vision + vision-language model pipeline routes each element to purpose-built models.
- **Multiple performance modes** — Toggle between performance modes optimized for speed, cost, or accuracy.
- **Handle every file** — 25 file types, 100+ languages, and 4 chunking strategies — all through one API.

### 02 / extract

Extract structured data from documents into any schema.

- **Scale to 1,000+ page files** — Smart chunking & merging strategies ensure accurate extraction without hitting token output or context limits.
- **Extract large tables** — Choose from multiple array extraction strategies, capable of extracting 1,000+ rows accurately across 100s of pages.
- **Precise bounding boxes** — Dedicated citation model generates granular citations for every extracted value, even inferred ones.
- **Low latency processing** — Fast mode extraction for real-time use cases.
- **Model-agnostic by design** — New foundation models are continuously benchmarked and integrated. Evaluate performance and upgrade versions safely.

### 03 / split

Segment multi-document files into individual subdocuments.

- **Large document splitting** — High-precision splitting that maintains accuracy on 2,000+ page files.
- **Instance detection** — Detect unique identifiers to separate multiple instances of the same document type, like splitting 50 invoices by invoice number.
- **Intelligent boundary handling** — Smart overlap handles mid-page boundaries while keeping context intact.
- **Cost-optimized splitting** — Toggle cost-optimized splitting for bulk jobs, or high precision for tough files.

### 04 / classify

Classify documents into pre-defined categories.

- **Optimized for cost** — Deploy fast and cheap classification at scale with purpose-built classifiers, without compromising on accuracy.
- **Memory** — Multimodal retrieval system that learns from past examples to handle edge cases where prompting falls short.
- **Real time classification** — Toggle low-latency classification to give instant feedback to your users.

### 05 / edit

Detect form fields and fill them programatically.

- **Comprehensive field support** — Detect and fill in checkboxes, signatures, text fields, tables, dropdowns, multi-line paragraphs, and character-per-box inputs (like SSNs).
- **Dynamic form filling for agents** — Use natural language instructions to map data to the right fields automatically, no template required.
- **Template-based filling** — For static forms, detect field positions once and fill them at scale deterministically.
- **Reliability at any scale** — Fast, reliable results even on complex forms with hundreds of fields.

## Benchmarks

### RealDoc-Bench

Document parsing: RealDoc-Bench **95.7%** #1 on Document Q&A

**State of the art parsing on the hardest production documents.**

RealDoc-Bench tests production documents where layout, reading order, and field relationships determine downstream answer quality.

Views: Document Q&A, Layout, Cost, Latency

#### Document Q&A

Field-level answer accuracy from parsed markdown across 1,359 prompts and 581 documents.

Vertical: Full benchmark, Finance, Real estate, Logistics, Healthcare (showing Full benchmark)

| System | % correct |
| --- | --- |
| Extend Performance Parse | 95.7% |
| Extend Light Parse | 90.5% |
| Gemini 3.5 Flash | 89.04% |
| Azure DI | 88.8% |
| AWS Textract | 70.5% |

#### Layout

Adjusted F1 against human-labeled layout regions across forms, tables, key-value areas, charts, and text blocks.

| System | Adjusted F1 |
| --- | --- |
| Extend Parse 2.0 | 0.847 |
| AWS Textract | 0.709 |
| Azure DI | 0.687 |
| PaddleOCR-VL | 0.684 |
| DotsOCR | 0.320 |

#### Cost

Cost per page plotted against full-benchmark Q&A accuracy.

| System | % correct | Cost per page |
| --- | --- | --- |
| Extend Performance Parse | 95.7% | $0.020 |
| Extend Light Parse | 90.5% | $0.00625 |
| Gemini 3.5 Flash | 89.4% | $0.0113 |
| Azure DI | 88.8% | $0.010 |
| AWS Textract | 70.5% | $0.015 |

#### Latency

Parser latency per page plotted against full-benchmark Q&A accuracy.

| System | % correct | Latency per page |
| --- | --- | --- |
| Extend Performance Parse | 95.7% | 19.67s |
| Gemini 3.5 Flash | 89.4% | 8.50s |
| Azure DI | 88.8% | 5.46s |
| AWS Textract | 70.5% | 6.63s |

[Read the benchmark](https://www.extend.ai/resources/realdocbench)

### PoliTax Split

Document splitting: PoliTax Split **+28.4 pts** largest F1 lift over raw model

**Extend's splitter closes the boundary-recall gap on long tax PDFs.**

PoliTax Split tests long compound tax PDFs where subtle boundaries cause raw frontier models to miss document breaks.

Views: F1 lift

#### F1 lift

Raw model F1 compared with the same model wrapped in Extend's splitting harness.

| System | F1 score |
| --- | --- |
| Claude Opus 4.6 | 72.48% |
| Gemini 3.1 Pro | 72.34% |
| GPT-5.4 | 72.31% |
| Gemini 3 Flash | 72.26% |
| Claude Opus 4.5 | 64.11% |

[Read the benchmark](https://www.extend.ai/resources/document-splitting-benchmark)

### Long Array Extraction Benchmark

Structured extraction: Long Array **99.2%** overall mean per-doc accuracy

**Accurate long-array extraction across complex production documents.**

LongArray-Extract tests whether systems can reconstruct large arrays from long documents with full cardinality.

Views: Accuracy, Speed, Completion

#### Accuracy

Mean per-document accuracy across all 45 PDFs.

| System | Chart label | % correct |
| --- | --- | --- |
| Extend MAX | Extend MAX | 99.2% |
| Claude Opus 4.7 | Opus 4.7 | 83.1% |
| Gemini Pro | Gemini Pro | 47.3% |
| GPT-5.5 | GPT-5.5 | 31.6% |
| Gemini 3.5 Flash | Gemini Flash | 31.1% |

#### Speed

Mean wall-clock latency plotted against aggregate mean per-document accuracy.

| System | % correct | Mean latency |
| --- | --- | --- |
| Extend MAX | 99.2% | 301s |
| Reducto Deep Extract | 97.4% | 846s |
| Claude Opus 4.7 | 83.1% | 532s |
| Gemini 3.1 Pro | 47.3% | 332s |
| LlamaParse Agentic | 34.7% | 149s |
| GPT-5.5 | 31.6% | 128s |

#### Completion

Document completion plotted against accuracy on completed documents.

| System | % correct | Documents completed |
| --- | --- | --- |
| Extend MAX | 99.2% | 100.0% |
| Reducto Deep Extract | 97.4% | 100.0% |
| Claude Opus 4.7 | 83.1% | 100.0% |
| Gemini 3.1 Pro | 47.3% | 100.0% |
| LlamaParse Agentic | 44.6% | 77.8% |
| GPT-5.5 | 31.6% | 100.0% |

[Read the benchmark](https://www.extend.ai/resources/long-array-extraction-benchmark)

## Production-ready document processing takes more than models.

A batteries-included toolkit to go from PDFs to production, fast.

- **Confidence scoring** — Flag uncertainty before production. Enable a multi-pass review agent that checks every output to flag potential errors, so you detect issues before your users do.
- **Fast mode** — Toggle between multiple processing modes. Low latency for real-time use cases, cost optimized for bulk jobs, or maximum accuracy when precision matters.
- **Composer Agent** — Skip manual prompt trial-and-error with an optimization agent. Upload examples and let Composer identify issues, automatically refine your schemas, and improve accuracy in the background.
- **Document workflows** — End-to-end orchestration for complex pipelines. Build multi-step workflows that parse, split, extract, validate, and route with versioning and durability out of the box.
- **Studio & evals** — Ditch the CLI scripts and empower your domain experts. Iterate on schemas, run evals, catch regressions, and ship with confidence, all from one intuitive interface.

## What leading AI teams say about Extend

Customers:

> Extend outperformed every solution we tested—other vendors, open source, and even foundation models. It now powers key document workflows across 30,000 customers, helping us build the most intelligent and modern financial platform out there.
>
> — Pedro Franceschi, CEO, Brex

[Read more](https://www.extend.ai/resources/how-brex-reached-99-accuracy-across-millions-of-financial-documents)

> We did a bakeoff, and Extend had the best results of any solution on the market. It eliminates an entire class of engineering problems around accuracy that we don't want to worry about.
>
> — Matt Hodgson, CTO, Vendr

[Read more](https://www.extend.ai/resources/vendr-unlocks-data-from-millions-of-documents-to-launch-new-products)

> We were able to replicate 6 months of work in 2 weeks (!) with Extend. We're now scaling this up across all 5 million people with cancer in our network, truly transforming our work against this disease.
>
> — George Ho, SR ML Scientist, Flatiron Health

> Extend sets the bar for what all vendors should be - a best in class product and a world class team. We couldn't be happier and are deeply grateful for our relationship with them.
>
> — Yan Lhert, Staff Engineer, Opendoor

> We tested AWS, foundation models, and newer vendors, and evaluated options on latency, accuracy, and the developer experience. Extend was the only one that could hit our performance bar, with the tooling we needed to scale.
>
> — Paul-Arthur Asselin, Senior Software Engineer, Mercury

[Read more](https://www.extend.ai/resources/mercury-case-study)

> Extend eliminates the ongoing maintenance cost of model tuning, scoring, evaluations, and more. We're able to focus on innovating on our core experience, instead of managing the infra.
>
> — Adam Litton, Staff Software Engineer, Checkr

[See all customer stories >](https://www.extend.ai/customers)

## Enterprise-grade security for your data.

Security:

- **Self-hosted deployment** — Run entirely on your infrastructure. Keep sensitive documents in-house with the same speed, accuracy, and features as our cloud.
- **SOC 2, HIPAA, & GDPR** — Built for regulated industries. Certified with regular third-party penetration testing, and trusted by F500s in every industry.