# I built a document extraction API that runs on regex (no LLM needed)

> Source: <https://dev.to/tuyentn23dot/i-built-a-document-extraction-api-that-runs-on-regex-no-llm-needed-3b0g>
> Published: 2026-09-22 04:44:18+00:00

I kept rewriting invoice parsing code on every freelance project. Same regex, same edge cases, same pain. So I made it a public API instead.

It's called **Invoice to JSON Extractor**. Four endpoints, one auth, text in, structured JSON out.

The interesting part: the base path doesn't use an LLM. It's a regex heuristic engine that handles ~70% of typical invoices in under 50ms, for zero ongoing cost.

| Endpoint | Input | Output | 
|---|---|---|
| `POST /v1/invoice/extract` | Invoice text | number, dates, vendor, VAT, totals, line items | 
| `POST /v1/receipt/extract` | Receipt text | store, date, items, total, payment method | 
| `POST /v1/resume/extract` | CV/resume text | name, email, phone, skills, experience | 
| `POST /v1/bank-statement/extract` | Bank statement text | account, transactions, running balance | 

Bilingual out of the box (English + Vietnamese).

When I started, I assumed I'd need an LLM for this. But most invoices follow one of maybe 15 common layouts. Regex handles these in 40-60ms with zero API calls.

The LLM is a **fallback**, not the primary engine. If the regex extractor returns a low confidence score, the API escalates to DeepSeek or Groq for that specific request.

Result: ~95% of requests never hit an LLM. Average cost per request rounds to zero.

Input:

```
INVOICE
Invoice #: INV-2024-001
Date: 15/03/2024
Vendor: Acme Supplies Ltd.
VAT: GB123456789
Bill To: Widget Corp

Widget A  10  5.00  50.00
Widget B  3   20.00 60.00

Subtotal: 110.00
VAT 20%: 22.00
Total: 132.00 GBP
```

Output:

```
{
  "success": true,
  "invoice": {
    "invoice_number": "INV-2024-001",
    "invoice_date": "2024-03-15",
    "currency": "GBP",
    "vendor_name": "Acme Supplies Ltd.",
    "vendor_tax_id": "GB123456789",
    "customer_name": "Widget Corp",
    "subtotal": 110.0,
    "tax_amount": 22.0,
    "total": 132.0,
    "line_items": [
      {"description": "Widget A", "quantity": 10.0, "unit_price": 5.0, "amount": 50.0},
      {"description": "Widget B", "quantity": 3.0, "unit_price": 20.0, "amount": 60.0}
    ],
    "confidence": 0.9
  },
  "model": "heuristic:v1",
  "processing_ms": 42
}
curl -X POST "https://invoice-to-json-extractor1.p.rapidapi.com/v1/invoice/extract" \
  -H "content-type: application/json" \
  -H "X-RapidAPI-Key: YOUR_KEY" \
  -H "X-RapidAPI-Host: invoice-to-json-extractor1.p.rapidapi.com" \
  -d '{"content":"Invoice #: INV-1\nDate: 2024-01-15\nTotal: 132.00 USD"}'
```

Get a free key here: [Invoice to JSON Extractor](https://rapidapi.com/tuyentn23/api/invoice-to-json-extractor1)

Free tier: 500 calls/month. Paid tiers start at $9/month for 2,000 calls.

If you build anything that processes invoices or receipts, I'd love feedback on what fields you actually need. The API is small and I can iterate fast.

*Stack: FastAPI + Pydantic, deployed on Render free tier, billing via RapidAPI. Total monthly cost so far: $0.*