Document archives to JSON or MD
Turn your whole document archive into JSON or MD
At a fraction of the AI bill
AI can easily convert PDF to JSON so what is the problem?
Your token bill is the problem.
What if I told you that you can convert your whole document library, with millions of documents in various formats, and save thousands on your AI bill?
{
"element_type": "table",
"self_ref": "#/tables/0",
"bounding_boxes": [
{
"page_number": 7,
"x0": 72,
"y0": 418,
"x1": 486,
"y1": 452
}
]
}
An illustrative cost comparison
UBS, Deutsche Bank and Merrill Lynch each publish between 40,000 and 50,000 research reports a year.
Let's use one of them as an example:
| Input / 1M tokens | Output / 1M tokens | Per 1,000 pages | Cost for 9.6 million pages | Saving in USD | |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.35 | $3,368 | -$1,432 |
| DeepSeek V4.1 Flash | $0.15 | $0.60 | $0.43 | $4,116 | -$684 |
| Mistral Small 4 | $0.15 | $0.60 | $0.43 | $4,116 | -$684 |
| docs2json | $0.50 | $4,800 | - | ||
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $1.04 | $9,979 | +$5,179 |
| Gemini 2.5 Flash Deprecated Dec 2026 | $0.30 | $2.50 | $1.70 | $16,343 | +$11,543 |
| GPT-6.1 Terra | $1.00 | $5.00 | $3.51 | $33,677 | +$28,877 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $6.24 | $59,875 | +$55,075 |
| GPT-5 Deprecated Dec 2026 | $1.25 | $10.00 | $6.82 | $65,496 | +$60,696 |
| Claude Sonnet 5 | $2.00 | $10.00 | $7.02 | $67,354 | +$62,554 |
| Gemini 3 Pro | $2.00 | $12.00 | $8.32 | $79,834 | +$75,034 |
| Claude Sonnet 4 Retired Jun 2026 | $3.00 | $15.00 | $10.52 | $101,030 | +$96,230 |
258 input and 650 output tokens per page, at standard API prices.
And that is just for a single run. Every schema change, backfill, failover environment and reprocessing repeats it.
When a provider retires the model your pipeline runs on, you pay twice: the migration, and then a higher price per page.
The same comparison at other archive sizes:
| 10 providers96M pages | 40 providers384M pages | Filings exhibits500M pages | 600 providers5.76B pages | |
|---|---|---|---|---|
| Claude Haiku 5.5 | $33.7k | $134.7k | $175.4k | $2.02M |
| DeepSeek V4.1 Flash | $41.2k | $164.6k | $214.3k | $2.47M |
| Mistral Small 4 | $41.2k | $164.6k | $214.3k | $2.47M |
| docs2json | $48k | $192k | $250k | $2.88M |
| Gemini 3.1 Flash-Lite | $99.8k | $399.2k | $519.8k | $5.99M |
| Gemini 2.5 Flash Deprecated Dec 2026 | $163.4k | $653.7k | $851.2k | $9.81M |
| GPT-6.1 Terra | $336.8k | $1.35M | $1.75M | $20.21M |
| Gemini 3.5 Flash | $598.8k | $2.40M | $3.12M | $35.93M |
| GPT-5 Deprecated Dec 2026 | $655k | $2.62M | $3.41M | $39.30M |
| Claude Sonnet 5 | $673.5k | $2.69M | $3.51M | $40.41M |
| Gemini 3 Pro | $798.3k | $3.19M | $4.16M | $47.90M |
| Claude Sonnet 4 Retired Jun 2026 | $1.01M | $4.04M | $5.26M | $60.62M |
Also converted: insurance claims, court filings, RFPs and DDQs, patent documents, and scanned archives.
Contact us if any of this is you:
You spend thousands on LLM runs
You pay for every output token, and again on every retry, test run or schema change.
You test with a cheaper model
A faithful staging run costs the same as production, so quality is only proven after the real run.
Your tables break in conversion
Multi-column pages and tables that cross a page break come out as loose text.
You have your own JSON schema
Someone writes code to map each output field to yours, and a change on either side breaks it.
You depend on one AI model
If it is retired, repriced or changed, you reconfigure, re-test quality and cost the replacements.
You need an on-prem deployment
No provider may keep the pages or train a model on them, so the pipeline has to run inside your network.
You cannot trace a value back
Nothing in the output points to the page, so checking a number means hunting through the file.
Your archive includes (low quality) scans
A faint page or a phone photo defeats the text layer, so extraction fails or comes back wrong.
What the JSON contains
Let's look at NVIDIA's 2027 second quarter earnings release, for example. Select a box on the page or a row in the list to see that element's JSON.
Elements in reading order
Source: NVIDIA Form 8-K, Exhibit 99.1, filed with the SEC on 26 August 2026
Also available
- On-prem deploymentYour documents never leave your servers.
- Metadata extractionFields read off the page for you, such as invoice numbers, counterparties or reporting periods.
- Your custom JSON schema exportThe document model mapped into your fields, delivered as a schema you can inspect and re-run.
- Need markdown instead of JSON?The same archive converts to Markdown for LLMs and RAG, with tables kept as tables and the page and box behind every block.