HyperAPI: one API for every enterprise document, plus ten open benchmarks to test it

Enterprises run on documents, and only some of them are invoices. A single AP inbox, claims queue or onboarding portal receives purchase orders, bank statements, bills of lading, KYC packs, payslips, contracts, lab reports and emails. They arrive as clean PDFs, phone photos, stamped scans and 40-page bundles. Most pipelines that handle them are built from separate OCR, classification, field-mapping, validation and redaction tools, often maintained per format or per client.

HyperAPI replaces that stack with five operations behind one endpoint: parse, extract, classify, split and redact. Each returns schema-validated JSON with a confidence score and a source bounding box for every field. It's the same document intelligence that runs inside Hyperbots' finance agents, where it handles more than 1M invoices a year for a single customer. Today we're also releasing ten fully synthetic datasets on the Hugging Face Hub, so you can test it, or any other document AI pipeline, without sharing a single real document.

Five operations

You upload a file once and get a document_key, valid for 24 hours, then run any operation against it. Every response carries request_id, model_used, duration and page count for audit. You can call the operations synchronously, asynchronously (X-Async: true) or in batches with webhooks. Results are cached for 24 hours, and cached re-runs aren't billed.

Get a free key (30 pages/month, no credit card), set HYPERAPI_KEY, and follow the quickstart to use the SDK or REST API:

from hyperapi import HyperAPIClient

with HyperAPIClient() as client:
    kind  = client.classify("packet.pdf")                 # what is it?
    parts = client.split("packet.pdf")                    # where does each document start/end?
    doc   = client.process("packet.pdf")                  # parse + extract
    safe  = client.redact("packet.pdf", mode="deidentify")

Any document, your schema

The built-in financial format covers invoices and receipts (213 pretrained invoice fields). For everything else, send the output you want as a JSON template, and HyperAPI fills it in:

curl -X POST "https://apis.hyperbots.com/v1/extract?category=non_financial" \
  -H "X-API-Key: $HYPERAPI_KEY" \
  -F "document_key=..." \
  -F 'schema={"policy_number":null,"claimant_name":null,"date_of_service":null,"amount_claimed":null}'

The same call works for a bill of lading, a KYC form, an offer letter or a maintenance report. Every row of our enterprise extraction dataset ships with the template to send.

Built for agents

Coding agents follow goals well but don't read unstructured documents reliably. HyperAPI's five operations can be exposed over MCP (available on request), so Claude Code, Cursor, GitHub Copilot, Codex, OpenCode or any MCP client can build the document pipeline a goal needs. With Agent Builder (beta), a request like "reconcile this month's vendor invoices against our open POs, flag anything over $500 and mask the bank details" runs as classify, then split, then extract, then two agent steps (match and flag), then redact. It returns one JSON ledger with a citation for every field.

What it's built on

HyperAPI runs on HyperLM, Hyperbots' mixture-of-experts document stack. It combines small, vision and large language models with multimodal encoders for layout, text and image, and decoders specialised by task and field. It's trained on 35M+ real-world documents and 0.5B synthetic samples, covers 100+ languages and is retrained continuously. The team's DocuLite paper (4th Deployable AI Workshop, AAAI 2026) shows how far synthetic data can take small extraction models, which is the same idea behind the datasets below.

Reported results from production (client names withheld):

• 1M+ invoices a year across 21 entities on SAP and NetSuite: 99%+ extraction accuracy, 80% less manual AP work and 40% lower AP cost.
• 300K+ invoices a year into SAP ECC for a $10B manufacturer: 40% fewer outsourced (KPO) staff.
• 1M+ dense directory pages in 10 days: column boundaries and reading order intact, and every listing turned into a record.
• One parse call across five formats (vector PDF, scanned PDF, XPS, Outlook email, Excel): 97% median text recall on the client's own files.

Ten open benchmarks

Real enterprise documents are hard to share, which makes document AI hard to evaluate in the open. So we generated our test sets. Every dataset below is fully synthetic, labelled by the generator that drew it (no annotation noise), formatted like HyperAPI's output and licensed CC BY 4.0. The generator is published alongside the data.

They're deliberately clean about what they are: synthetic, English, nine locales. Use them for regression tests, model comparisons and pipeline development, then confirm on your own documents.

Try it

• 🧾 HyperAPI on Hugging Face and the collection
• 🔑 Free API key (30 pages/month) and a no-sign-up sandbox
• 📚 Docs and the Python SDK
• 🏢 Enterprise: SOC 1 Type 2, SOC 2 Type 2, ISO 27001, BYOK, IP allow-listing and SSO. Evaluations run as a POC on your own documents, scored field by field. Contact hyperapi-support@hyperbots.com.