	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Gemini API for Data Extraction and Transformation

Gemini API for Data Extraction and Transformation

Author: Venkata Sudhakar

Data extraction and transformation is one of the highest-value use cases for LLMs in enterprise settings. The Gemini API can parse invoices, purchase orders, web pages, and email attachments to extract structured fields - with far higher accuracy than traditional regex or template-based parsers, especially when the format varies across suppliers or sources.

ShopMax India receives supplier invoices in dozens of different formats - some PDF, some scanned, some in tables, some in free text. The finance team previously spent hours manually entering invoice data into their ERP system. With Gemini, they extract and transform invoice data automatically at over 95% accuracy, reducing manual review to edge cases only.

The below example shows how to extract invoice data from unstructured text and transform it into a structured format ready for database insertion.

import vertexai
from vertexai.generative_models import GenerativeModel
import json

vertexai.init(project="shopmax-india", location="asia-south1")
model = GenerativeModel("gemini-2.0-flash")

INVOICE_SCHEMA = {
    "invoice_number": "string",
    "invoice_date": "YYYY-MM-DD",
    "supplier_name": "string",
    "supplier_gstin": "string",
    "line_items": [
        {"description": "string", "qty": "number", "unit_price": "number", "total": "number"}
    ],
    "subtotal": "number",
    "gst_amount": "number",
    "total_amount": "number",
    "payment_terms": "string"
}

def extract_invoice(raw_text: str) -> dict:
    prompt = f"""Extract invoice data from the text below.
Return ONLY valid JSON matching this schema:
{json.dumps(INVOICE_SCHEMA, indent=2)}

Use null for missing fields. Convert all amounts to numbers (no commas or currency symbols).

Invoice text:
{raw_text}

JSON:
"""

response = model.generate_content(
        prompt,
        generation_config={"temperature": 0.0, "response_mime_type": "application/json"}
    )
    return json.loads(response.text)

# Sample invoice text from a ShopMax India supplier
invoice_text = """
TAX INVOICE
Samsung Electronics India Pvt Ltd
GSTIN: 29AABCS1429B1ZB
Invoice No: SI-2024-48291  Date: 18/03/2024

To: ShopMax India, Koregaon Park, Pune - 411001

Galaxy S24 Ultra 256GB    Qty:10  Rate: Rs.1,12,999  Amount: Rs.11,29,990
Galaxy Buds3 Pro          Qty:25  Rate: Rs.17,990    Amount: Rs.4,49,750

Subtotal: Rs.15,79,740
GST 18%: Rs.2,84,353
Grand Total: Rs.18,64,093
Payment: Net 30 days
"""

data = extract_invoice(invoice_text)
print(f"Invoice: {data['invoice_number']}")
print(f"Supplier: {data['supplier_name']}")
print(f"Total: Rs {data['total_amount']:,.0f}")
print(f"Line items: {len(data['line_items'])}")

It gives the following output,

Invoice: SI-2024-48291
Supplier: Samsung Electronics India Pvt Ltd
Total: Rs 1,864,093
Line items: 2

The below example shows how to validate the extracted data and insert it into a database after transformation.

def validate_and_transform(invoice: dict) -> dict:
    # Validate required fields
    required = ["invoice_number", "supplier_name", "total_amount", "line_items"]
    missing = [f for f in required if not invoice.get(f)]
    if missing:
        raise ValueError(f"Missing required fields: {missing}")

# Validate GST calculation (18% standard rate)
    if invoice["subtotal"] and invoice["gst_amount"]:
        expected_gst = round(invoice["subtotal"] * 0.18, 2)
        actual_gst = invoice["gst_amount"]
        if abs(expected_gst - actual_gst) > 10:
            invoice["gst_validation"] = "MISMATCH - manual review needed"
        else:
            invoice["gst_validation"] = "OK"

# Transform for ERP insertion
    return {
        "erp_invoice_id": invoice["invoice_number"],
        "vendor_gstin": invoice.get("supplier_gstin"),
        "amount_excl_tax": invoice["subtotal"],
        "tax_amount": invoice["gst_amount"],
        "amount_incl_tax": invoice["total_amount"],
        "line_count": len(invoice["line_items"]),
        "status": "pending_approval",
        "validation": invoice.get("gst_validation", "OK")
    }

erp_record = validate_and_transform(data)
for k, v in erp_record.items():
    print(f"{k}: {v}")

It gives the following output,

erp_invoice_id: SI-2024-48291
vendor_gstin: 29AABCS1429B1ZB
amount_excl_tax: 1579740
tax_amount: 284353
amount_incl_tax: 1864093
line_count: 2
status: pending_approval
validation: OK

The combination of Gemini extraction and custom validation logic handles ShopMax India supplier invoices regardless of format variation. By setting temperature to 0.0 and using response_mime_type application/json, the extraction is deterministic and directly parseable - eliminating the brittle string parsing that traditionally makes invoice automation fragile in production.

Send your comments, suggestions or queries regarding this site to [email protected].