Refractr

What is Refractr?

Refractr is a developer API that pulls structured data out of messy text, emails, PDFs, and photos. You send a document with your own JSON template and get validated JSON back, with a confidence score per field. It runs a custom extraction model and answers in about half a second.

Top Features:

• Schema per call: send any JSON template with each request, with no setup step.
• Built-in OCR: reads scanned files, phone photos, and screenshots when switched on.
• Confidence scores: each field returns a score so weak values can go to review.

Use Cases:

• Invoice processing: capture vendors, totals, dates, and line items from supplier invoices.
• Logistics paperwork: read bills of lading and delivery notes, including handwritten names.
• Ticket triage: turn incoming support emails into clean fields for routing and tracking.

Who Can Use Refractr?

• Developers: add document parsing to any app with a single REST call.
• Finance and operations teams: automate manual data entry from supplier invoices and shipping documents.
• Support teams: structure customer emails so tickets get sorted without manual reading.

Pricing

• Alpha allowance (free): 100 extractions every day while the product stays in alpha.
• Credits (€1 per 200): one credit per extraction, with a five euro minimum top-up.
• OCR extractions (2 credits each): scanned files and photos cost double, failed calls cost nothing.

Pros and Cons

Pros:

• Fast responses: median latency sits near half a second, far quicker than large models.
• Clean output: benchmarks show zero invalid JSON responses across its test set.
• Fair billing: flat per-extraction pricing applies regardless of schema size or complexity.

Cons:

• Early stage: the product is still in alpha and marked under construction.
• Simple fields only: works best with about ten concrete fields, not summaries or sentiment.
• No batching: the docs mention no batch endpoints, webhooks, or official SDKs.

FAQs:

1) What file types does it accept?

PDF, DOCX, ODT, common image formats, TIFF scans, and plain text or email.

2) Is it just a wrapper around GPT?

No, it runs its own custom-trained model instead of passing requests to OpenAI.

3) How accurate is it?

It scores 82 out of 100 on its benchmark, close to slower models.

4) Are there rate limits?

Each API key can send up to 60 requests per minute right now.

5) Can I process documents asynchronously?

Yes, set wait to false and poll the job endpoint for results.