A billion-dollar industry hiding behind a file extension.
PDF parsing is not reported as a standalone market, so the headline needs precision. But the ecosystem built around making PDFs useful—from document software to AI ingestion infrastructure—is unquestionably worth billions.
Market signals
The category is already being funded, priced, and scaled.
$3.18B Adobe Document Cloud FY2024 revenue
$3.48B Adobe Document Cloud ending ARR
$27.5M Total capital raised by LlamaIndex
3M+ Monthly LlamaIndex downloads
5M+ Unstructured open-source downloads
$0.03 Unstructured price per page
1B+ Pages processed by Reducto
$108M Funding reported by Reducto
Inside this story
01 PDFs preserve appearance—not understanding.02 Correctness, speed, and good APIs rarely arrive together.03 Make the engine itself worthy of the workflow.
01 The format
PDFs preserve appearance—not understanding.
document.pdf
appearance ≠ meaning
PDF 2.0 runs 1,003 pages and has no reference implementation. The specification defines valid files, but it does not tell implementers how to recover malformed ones. Every parser must decide when to reject, repair, or make an educated guess.
The format also lacks a dependable semantic layer. Reading order, paragraphs, headers, tables, and even spaces may need to be reconstructed from fonts, drawing instructions, and coordinates.
02 The gap
Correctness, speed, and good APIs rarely arrive together.
parser.matrix
parser
correctness
speed
API / DX
integrity correctness · reliability
engine performance
product UX · usability
interface DX · ergonomics
core
pick two? designed together.
Core PDF starts with a refusal: we are not willing to trade correctness for speed, reliability for usability, or performance for an ergonomic API. Those qualities have to be designed as one system.
Existing parsers are not universally bad; they make different trade-offs. Some focus on editing, some on text, some on rendering, and some on AI-ready output. Coverage can stop at difficult fonts, scanned pages, nested tables, unusual cross-references, or damaged files.
At scale, the trade-offs become product problems: slow ingestion, memory spikes, silent omissions, shuffled reading order, and layers of fallback code. Developers then inherit a fragmented API surface and must translate low-level parser concepts into the simple outcomes their applications actually need.
The market tells on itself: LlamaIndex built LlamaParse for complex documents, Unstructured sells transformation by the page, and Reducto has raised heavily around document intelligence. These companies validate the need—even as they pursue a different, primarily hosted and AI-assisted layer of the stack.
03 Our bet
Make the engine itself worthy of the workflow.
core-pdf
01 bytes 02 parse 03 layout 04 output TXT MD ▦ IMG
Core PDF is our effort to make working with PDFs feel straightforward. It should understand messy real-world documents, return useful results quickly, and give developers an API that feels predictable from the first call.
We are building it around outcomes people can verify: whether it handles difficult PDFs correctly, how fast it runs, and how easy it is to use. We will publish benchmarks and compatibility results as it matures.
Reliable results, even with messy PDFs Fast enough for demanding workloads A clear API developers enjoy using
