Accuracy
84.3Index, out of 100
Field-level accuracy · Higher is better
Structured extraction API
Bend raw data into structure.
Refractr turns chaotic payloads into validated output fields a computer can understand, with a single API call and your own JSON schema.
Why Refractr
Small, fast, honest.
Latency
~500 ms per call
A custom-trained extraction model, not a general-purpose LLM wrapped in a prompt. Predictable sub-second response times, not multi-second chains.
Schema, your terms
Zero compile time
Pass any JSON schema on every call and get a result immediately. No pre-compilation, no warm-up step, no per-schema deployment. Change fields between requests with no penalty.
Not an LLM wrapper
Purpose-built model
Refractr runs a custom, optimized model hosted by us, not an OpenAI or Anthropic passthrough. One extraction task, done faster and cheaper than general-purpose models.
Benchmarks
Measured, not claimed.
Speed
0.49Seconds, median
p50 seconds per document · Lower is better
Reliability
0.0%Invalid JSON
Responses that fail to parse · Lower is better
Speed vs accuracy
8×Faster than any more accurate system
Accuracy index vs median latency per document · Top left is better
Scroll sideways to see the whole chart.
OCR speed
<0.8Seconds per page
Scan or photo in, fields out · Lower is better
Accuracy index is field-level accuracy averaged across several document test sets. Latency is end to end per document. The most attractive quadrant is the top-left quarter of the chart: under 5 s and an accuracy index of 70 or more. OCR speed is per page, from scan or photo to extracted fields, with all systems measured by us on the same documents and Refractr shown at its slowest measured case. Extraction benchmark run 28 Sep 2026.
Use cases
Same API, anything it reads.
Market news → tickers
BREAKING: Nvidia soars 12% to ~$187 after crushing Q4 earnings. Revenue hit $22.1B vs $20.4B expected.
↓
{
"ticker": "NVDA",
"move_pct": 12.0,
"revenue_billions": 22.1,
"currency": "USD"
}
Invoice → line items
Invoice #2847
From: Acme Corp
Date: 2026-01-15
Total: €1,249.00 VAT 21%
↓
{
"invoice_no": "2847",
"vendor": "Acme Corp",
"date": "2026-01-15",
"total_eur": 1249.00
}
Email → ticket fields
Hi, order #88420 arrived damaged, box was crushed. Refund would be great. Mara, [email protected]
↓
{
"order_id": "88420",
"issue": "damaged",
"action": "refund",
"contact": "[email protected]"
}
Pricing
One credit per extraction.
Pay-as-you-go
1 credit / extraction
€1 = 200 credits · top up from €5
- 100 free extractions per day on every account
- Failed extractions are free
- Scans and photos with OCR: 2 credits
- Flat pricing
- Bring your own JSON schema on every call
- Top up any amount via Stripe
Common questions
Before you sign up.
Is Refractr an LLM wrapper?
No. Refractr runs a custom-trained, optimized extraction model hosted by us, not a passthrough to OpenAI, Anthropic, or any general-purpose LLM. It's purpose-built for one task: filling a JSON schema from raw text, which is why it's faster than prompt-based approaches.
What happens if a field is missing from the input?
Fields declared with typed placeholders (like "__date__") are guaranteed to come back as that type or null, never an empty string or a mistyped value. That guarantee covers the shape of the response; extraction is model-based, so validate values against your own rules and enforce required fields in your own schema validator downstream.
Is my data used to train anything?
Not unless you choose to. By default, documents and results are kept for 30 days for debugging and support, then deleted automatically. You can opt in to help train the model when you sign up or from your dashboard, which raises your free allowance from 20 to 100 extractions per day, and switch it off at any time. Your data is never sold or shared.
How does pricing work?
One credit per successful extraction: flat, regardless of document or schema size. Failed extractions (schema violations, timeouts) are free. Every account gets 100 free extractions per day, refreshed daily. For higher volume, top up any EUR amount via Stripe.
Do I need to register or compile a schema first?
No. Pass the schema on every call. There's no compile step, no warm-up, no per-schema deployment. Change fields between requests with no penalty.
Ship structured data today.
Five lines of code, five minutes to first response.