Verification & Self-Correction

Strutex v0.8.0 introduces Verification, a powerful mechanism to improve extraction accuracy by using the LLM to audit and correct its own work.

Overview

Extraction errors can happen, especially with complex documents. The Verification feature runs a two-step workflow:

1. Extraction: The document is processed normally to generate structured JSON.
2. Audit: The result from step 1 is fed back to the LLM (or a different model) along with the original document. The model acts as a "strict auditor" to check for discrepancies and fix them.

This "Reflexion" pattern significantly reduces hallucinations and missing fields.

---

Usage

Automatic Verification

The simplest way to use verification is to pass verify=True to the process (or aprocess) method.

from strutex import DocumentProcessor
from strutex.schemas import INVOICE_US

processor = DocumentProcessor()

result = processor.process(
    file_path="invoice.pdf",
    prompt="Extract invoice data",
    schema=INVOICE_US,
    verify=True  # 👈 Enables self-correction loop
)

Manual Verification

You can also run verification on a result you already have (e.g., from a previous run or a different system).

previous_result = {
    "invoice_number": "INV-001",
    "total": 500.00  # Suspicious value
}

verified_result = processor.verify(
    file_path="invoice.pdf",
    result=previous_result,
    schema=INVOICE_US
)

Async Support

Verification is fully supported in async workflows:

result = await processor.aprocess(
    ...,
    verify=True
)

# Or manually:
result = await processor.averify(...)

---

Configuration

Custom Verification Prompt

By default, Strutex uses a generic "strict auditor" system prompt. You can customize this to look for specific errors.

result = processor.process(
    ...,
    verify=True,
    verify_prompt="Check strictly that the 'total' matches the sum of line items."
)

---

Performance & Cost

Enabling verification doubles the number of LLM calls (1 extraction + 1 verification). This increases:

• Latency: Expect roughly 2x processing time.
• Cost: Expect roughly 2x token usage.

Usage Tracking

With verify=True, usage statistics (tokens/cost) in the returned result only reflect the verification pass. If you need total cost tracking, ensure you are logging usage from your quota manager or provider dashboard.

Best Practices

• Use verify=True for critical data fields where accuracy is paramount (e.g. financial totals).
• For high-volume low-value documents, you might skip verification to save costs.
• Consider using a faster/cheaper model for extraction and a stronger model for verification (currently requires manual 2-step calls).