Adding Validation¶

Learn to validate extracted data for accuracy and consistency.

Why Validate?¶

LLMs can make mistakes:

Math errors — totals don't add up
Date formatting — inconsistent formats
Missing required fields — schema not enforced
Hallucinations — data invented from nothing

Validation catches these before they reach your application.

Built-in Validators¶

SchemaValidator (Always Active)¶

Validates that output matches your schema structure:

from strutex import DocumentProcessor
from pydantic import BaseModel

class Invoice(BaseModel):
    invoice_number: str
    total: float

processor = DocumentProcessor(provider="gemini")
result = processor.process("invoice.pdf", "Extract", model=Invoice)
# Schema validation happens automatically

SumValidator¶

Checks that line items add up to totals:

from strutex import DocumentProcessor
from strutex.validators import ValidationChain, SchemaValidator, SumValidator

# Create validation chain
validators = ValidationChain([
    SchemaValidator(),
    SumValidator()  # Checks: sum(items.price) == total
])

# The SumValidator will flag if totals don't match
# useful for invoices, receipts, purchase orders

DateValidator¶

Ensures dates are valid and consistent:

from strutex.validators import DateValidator

validators = ValidationChain([
    SchemaValidator(),
    DateValidator()  # Checks date format validity
])

Verification Loop (Self-Correction)¶

Let the LLM audit its own work:

result = processor.process(
    file_path="invoice.pdf",
    prompt="Extract invoice",
    model=Invoice,
    verify=True  #  LLM reviews and corrects errors
)

How it works:

First pass: Extract data
Second pass: LLM reviews extraction against document
Returns corrected result if errors found

This catches ~30% more errors at 2x API cost.

Custom Validation¶

Add your own business logic:

from strutex import DocumentProcessor

def validate_invoice(result, context):
    errors = []

    # Check PO number format
    if not result.get("po_number", "").startswith("PO-"):
        errors.append("PO number must start with 'PO-'")

    # Check date is not in future
    from datetime import datetime
    if result.get("date"):
        inv_date = datetime.strptime(result["date"], "%Y-%m-%d")
        if inv_date > datetime.now():
            errors.append("Invoice date cannot be in the future")

    if errors:
        result["_validation_warnings"] = errors

    return result

# Register as post-process hook
processor = DocumentProcessor(
    provider="gemini",
    on_post_process=validate_invoice
)

Validation Strategies¶

Scenario	Strategy
Invoices/receipts	`SumValidator` + `verify=True`
Dates matter	`DateValidator`
Critical data	All validators + custom rules
High volume	`SchemaValidator` only (fast)
Compliance	Custom hooks for business rules

Error Handling¶

from strutex.validators import ValidationChain, SchemaValidator, SumValidator

chain = ValidationChain([SchemaValidator(), SumValidator()])

# Validate manually
is_valid, errors = chain.validate(result, schema)

if not is_valid:
    print(f"Validation failed: {errors}")
    # Handle errors...

Next Steps¶

Want to...	Go to...
Learn about caching	Caching
Add security checks	Security Layer
Use in RAG pipelines	Integrations