How Extend Is Redesigning Document Parsing for AI Agents

Extend released Parse 2.0 in May 2026, along with something called RealDoc-Bench — a testing tool for measuring how well document parsers handle real-world files. The dual launch comes about a year after the company raised $17 million in funding from Innovation Endeavors (the venture firm co-founded by Eric Schmidt), Y Combinator, and other investors.

These announcements signal a shift in how companies are approaching a stubborn engineering problem: extracting structured information from documents like forms, contracts, and invoices in ways that make sense to artificial intelligence systems.

What Parse 2.0 Does Differently

The core innovation in Parse 2.0 is called "layout-first" parsing. Here's why that matters.

Most document parsing tools work by reading text left-to-right, top-to-bottom — the way humans read a page — and then trying to figure out what the layout was. If a financial statement has numbers in three columns side by side, a traditional text-first parser will squash all three columns into a single stream of words. The AI agent reading that output then has to guess which numbers belonged where.

Parse 2.0 does the opposite: it preserves the spatial layout and structure from the beginning. It treats the physical position, bounding boxes, and hierarchical relationships as core data, not an afterthought. Anyone who has tried to extract data from multi-column financial tables or nested forms knows why this matters — it's the difference between getting the structure right and having to reconstruct it downstream.

The API is built explicitly for agent consumption, not for human reading. That distinction shapes the output: instead of trying to produce something that looks clean in a PDF viewer, it emits the structured metadata that orchestration systems (particularly AI agents) actually need — precise coordinates, semantic block types, bounded regions.

Extend's open-source UI library, called Extend UI, hints at what this output looks like. The library provides visual components for PDF and spreadsheet viewing, file upload, and notably, something called "bounding box citations" — a way to highlight exactly where in the original document an extracted fact came from. That precision is only possible if the API is emitting coordinate data accurate enough to pinpoint text location.

RealDoc-Bench: The Problem It Solves

For decades, document parsing has been tested using datasets of clean, nicely formatted PDFs — mostly single-column text, usually in English, carefully curated for academic research. Real enterprise documents look nothing like that.

In the real world, you encounter scanned forms with skew and rotation artifacts, spreadsheets exported to PDF, multi-language invoices, and contracts that have been printed, signed, scanned, and compressed multiple times. Traditional benchmarks miss all of this because they're not measuring against documents that actually exist in production.

RealDoc-Bench is Extend's answer: a benchmark built from actual enterprise documents, not synthetic or academic ones. Extend published it publicly alongside Parse 2.0, which is an unusual move. Most vendors prefer to keep their own benchmarks proprietary — publishing the standard by which your product will be measured invites external scrutiny.

We have seen this pattern before, at a different layer of the technology stack. In the mid-2010s, cloud and CDN companies began publishing their own performance benchmarks, partly to build trust and partly because existing third-party measurements weren't capturing what mattered for modern workloads. Some of those became industry standards once independent researchers validated them. Whether RealDoc-Bench follows that arc depends on whether competitors and outside researchers find the methodology rigorous enough to adopt and rely on.

At minimum, making the benchmark public forces a more honest conversation about what "parsing quality" actually means when documents are messy and diverse.

The Open-Source Component Layer

Beyond the core API, Extend released Extend UI as an open-source library on GitHub. It includes 14 components covering the full workflow: PDF, DOCX, and XLSX viewers, file upload, page thumbnails, OCR overlays, those bounding box citations, and human review interfaces.

There is a practical strategy at work here. By giving away the front-end components, Extend reduces the friction for new customers to try the service — every engineering team that uses Extend UI is slightly less likely to switch to a competitor, simply because ripping out integrated components costs engineering time. It is a familiar playbook: commoditise the easy, surface-level parts to lower the barrier to adoption.

The human review interface component is worth noting. In any real-world document processing system, extractions that the parser is unsure about get routed to a human for verification. Providing an open-source component for that human-in-the-loop workflow signals that Extend is thinking about the full operational reality, not just the scenario where the parser works perfectly.

Funding and Capital Efficiency

The $17 million round was disclosed in mid-2025 and led by Innovation Endeavors, with Y Combinator's participation. Relative to some larger document AI competitors, it is a modest amount — suggesting either a tighter product scope or a deliberate choice to keep burn low. The API-first, agent-focused approach avoids the broader platform sprawl that typically drives higher headcount and spending.

Why This Matters Now

The document parsing problem itself is decades old — OCR technology dates to the 1970s. What has changed is the consumer: historically, parsed output went to rule-based systems that could accept messy, inconsistent output as long as key fields got extracted correctly. Now, AI agents and language models are consuming parsed documents.

Language models are more sensitive to structural coherence. A garbled table or an accidental sentence break in the parsed output can confuse downstream reasoning in ways that are hard to catch. That sensitivity is what makes layout-first parsing urgent now where it was nice-to-have five years ago. Extend is not alone in recognising this shift — other document intelligence platforms have made similar architectural moves — but the combination of an agent-oriented API, a real-world benchmark, and an open-source UI layer forms a fairly coherent strategy.

The practical test for any team evaluating document parsing tools is straightforward: does RealDoc-Bench's document distribution match your own? That is a question you can now answer by checking the benchmark yourself, rather than relying on a vendor's claims.

For engineering teams building document processing pipelines, tighter integration between document ingestion and agentic reasoning reduces one of the more persistent sources of brittleness in production workflows — a tangible gain in reliability.