AI-ready content

Your entire PDF corpus, ready for AI.

Scanned, typewritten, or born digital, Cashmere rebuilds every PDF as structured content AI can read, cite, and credit.

How we got here

Your catalog was built for people to find. Now AI has to understand it.

Most publishers hold thousands, often millions, of PDFs, laid down in layers over thirty years. Each layer solved the problem of its day.

1990s

You went digital

The paper archive was scanned into PDFs, one picture per page. The pages survived. What was on them, to software, did not.

Old or new, your PDFs weren't made for AI.

A scan from 1960 and a PDF from last month fail in different ways, but both fail. Getting your corpus ready means more than better OCR: every page has to be rebuilt as a document an AI can navigate, quote, and credit. That is what Cashmere does, for every PDF you hold.

The problem

Not all PDFs are created equal

Whether software can read a PDF depends on how it was made, and the oldest files, often the most valuable in a catalog, are the hardest of all. Here is one equation from each kind, three ways.

1
Hard
Born digital
LIGO 2016 · Kakeya 2025 · Beilstein 2026
The text is real, but it is stored as letters placed on a page, in the font’s own private encoding. Fractions fall apart, = becomes ¼, and a two-column page reads straight across both columns.
The page
What the PDF's own text says
```
M ¼ ðm1m2Þ3=5 ðm1 þ m2Þ1=5 ¼ c3 G 5 96 π−8=3f−11=3 _f 3=5 ;
```
Text layer: present, scrambled
What the omnipub says
$M = \frac{( m _{1} m _{2} ) ^{3/5}}{( m _{1} + m _{2} ) ^{1/5}} = \frac{c ^{3}}{G} [\frac{5}{96} π^{- 8/3} f^{- 11/3} f ˙]^{3/5},$
LIGO, 2016, page 3
2
Harder
Scanned print
Einstein 1905 · NACA 1953
A photograph of a printed page. The only text is an OCR layer added later, which reads ordinary words and gives up on everything else: equations, tables, and symbols.
The page
What the PDF's own text says
```
= 
Ot 
Or 
9: 
oder 
let.
```
OCR layer: fragments
What the omnipub says
$\frac{1}{2} (\frac{1}{V - v} + \frac{1}{V + v}) \frac{\partial τ}{\partial t} = \frac{\partial τ}{\partial x ^{'}} + \frac{1}{V - v} \frac{\partial τ}{\partial t},$
$\frac{\partial τ}{\partial x ^{'}} + \frac{v}{V ^{2} - v ^{2}} \frac{\partial τ}{\partial t} = 0.$
Einstein, 1905, page 9
3
Hardest
Typewritten originals
NASA TN D-233, Katherine Johnson, 1960
Typed, copied, and scanned from the copy. Uneven strikes, stamps, and margin notes, and Greek letters the OCR never learned: every λ, ω, and θ below is lost or misread.
The page
What the PDF's own text says
```
2_i_2e = (h2 + r_oET) + _E[t(e2e) -t(el)] - h I
```
OCR layer: symbols lost
What the omnipub says
$Δ λ_{1 - 2 e} = (λ_{2} + n ω_{E} T) + ω_{E} [t (θ_{2 e}) - t (θ_{1})] - λ_{1}$
NASA TN D-233, 1960, page 20

Why it matters

AI can only use what it can read.

A paper that fails ingestion is invisible: it can’t be searched, retrieved, cited, or licensed.

Wrong is worse than missing.

A misread number in a table or a garbled equation doesn’t look broken to an AI. It gets quoted, confidently.

The back catalog is the hard part.

Decades of scanned journals, reports, and books hold a publisher’s most distinctive content, and they are the nastiest PDFs of all.

How it works

From a page of ink to a document AI can use

One page of Katherine Johnson’s 1960 NASA paper, all the way through Cashmere. Every picture below is the real page and the real output.

STEP 01
Start with a PDF
A PDF records where ink goes on a page. To software, this page is a picture: the equations aren't text at all, and its OCR layer reads them as noise.
STEP 02
Read the page like a person
The Pro parser finds the structure a reader sees: which marks are a heading, a paragraph, an equation, a table, or a figure, and the order they're read in.
heading26
paragraph235
equation84
table11
figure13
list item24
787 blocks in the whole paper
STEP 03
Make every piece a block
Each piece becomes an omnipub block with a type and a stable address, so it can be searched, quoted, cited, and licensed on its own. Equations become real math.
Contents
1. SUMMARY3
2. INTRODUCTION4
3. SYMBOLS102
4. SATELLITE POSITION RELATIONSHIPS5
5. EXPRESSION FOR PLACING A SATELLITE OVER A SELECTEDPOSITION27
6. APPENDIX B1
7. EQUATIONS FOR CALCULATING AZIMUTH ANGLE, OBLATENESS EFFECTS,AND SATELLITE POSITION AND ILLUSTRATIVE EXAMPLES1
8. Azimuth Angle9
9. Oblateness Effects6
10. Satellite Position
11. Illustrative Examples of Orbit Calculations
28 sections · 787 blocks · 7,545 words · 13,716 tokens
STEP 04
Assemble the whole paper
The blocks are put back together as the paper itself: sections inside sections, in reading order, each with its own address. The result is one document an AI can navigate and measure, from its table of contents down to a single equation.

{
  "path": "dete/expr/ab/equa/azim/m1",
  "type": "math",
  "alttext": "[t(\\theta_{2e}) - t(\\theta_1)] \\approx \\frac{T}{360} (\\lambda_{2e} - \\lambda_1)",
  "fingerprint": "sha256:a73139c4833c122…",
  "section": "Azimuth Angle",
  "origin": {
    "page": 20,
    "region": [
      235.5,
      491.4,
      368.8,
      510.1
    ]
  }
}

STEP 05

Keep the link to the page

Every block remembers where it sat on the original page. An answer can cite the exact equation it used, and anyone can check it against the source.

Pick a paper below to see every page and every block side by side.

Why Cashmere

Parsed isn’t the same as AI-ready

Plenty of tools can pull text out of a PDF. An AI needs more: what each piece is, where it belongs, where it came from, and what it may be used for. That is the work Cashmere does, on the files other pipelines give up on.

Right on the hard pages

Layout, text, math, and tables are read from the page image itself, not from the broken text layer, so a 1905 scan comes out as clean as a 2025 preprint.

Structured like the work

Headings, sections, and reading order are kept, and every block gets a stable address. An AI can find the right passage instead of guessing from a wall of text.

Traceable to the source

Every block keeps its place on the original page. An answer can cite the exact paragraph or equation it used, and anyone can check it.

Licensed and measured

Rights travel with the content, and use is counted block by block, so you know what AI systems read and you get paid for it.

6 papers, 335 pages, from a 1905 scan to a 2025 preprint, converted by one pipeline with no hand corrections. Bring us your hardest PDFs.

Talk to us about your catalog

See for yourself

Six real papers. Open any one.

Each opens full screen: the original PDF on the left, the omnipub beside it, and the data behind any block you click.

Converted by Cashmere’s Pro parser, November 2026, and shown without hand corrections.

Thousands of PDFs or millions, ready for AI.

Cashmere converts your entire corpus quickly and accurately, from decades-old scans to last month’s releases, so AI can read it, cite it, and credit you.

Talk to us about your PDFs

Your entire PDF corpus, ready for AI.

Your catalog was built for people to find. Now AI has to understand it.

You went digital

Old or new, your PDFs weren't made for AI.

Not all PDFs are created equal

Born digital

Scanned print

Typewritten originals

Why it matters

AI can only use what it can read.

Wrong is worse than missing.

The back catalog is the hard part.

From a page of ink to a document AI can use

Start with a PDF

Read the page like a person

Make every piece a block

Assemble the whole paper

Keep the link to the page

Parsed isn’t the same as AI-ready

Right on the hard pages

Structured like the work

Traceable to the source

Licensed and measured

Six real papers. Open any one.

Katherine Johnson’s orbit paper

Compressible-flow tables

Gravitational waves detected

Terpenoid ring reactions

The Kakeya conjecture, solved

Special relativity, in German

Thousands of PDFs or millions, ready for AI.