# AI Data Extraction: Benchmarking GPT-4o vs. Veryfi

GPT-4o vs Veryfi for extracting financial data from irregular scans into Excel with Python: accuracy and cost per extraction across document types.

Benchmark · 12 Feb 2025, by dave8172. Page: https://quirkyagents.com/wizard/projects/pdf-excel-ai/

## Overview

Benchmarked across multiple document types. Identified cost-per-extraction and accuracy tradeoffs between both approaches.

![PDF to Excel using AI](https://quirkyagents.com/wizard/projects/pdf-excel-ai/pdf-excel-ai-preview.png)

This project addresses the challenge of "dark data" trapped in irregular, low-quality scans and PDFs. By building a dual-engine 
extraction system, I compared ****OpenAI’s GPT-4o**** (a general-purpose Multimodal LLM) and ****Veryfi**** 
(a specialized financial OCR API). 

The goal was to evaluate which system better handles real-world "noise" like faded text, line-item multipliers, and 
complex document layouts for downstream financial analysis.

---

## **Demo (Video)**

---

<video
  controls
  preload="metadata"
  poster="https://quirkyagents.com/wizard/projects/pdf-excel-ai/pdf-excel-ai-preview.png"
  style="width:100%; height: auto; aspect-ratio: 16 / 9; border-radius:8px;"
>
  <source
    src="https://pub-21da2bddd9cf4685b4374f3e9115db36.r2.dev/pdf-excel-ai.mp4"
    type="video/mp4"
  />
  Your browser does not support the vsideo tag.
</video>

---

## Tech Stack

* **Language:** Python 3.11+
* ****AI/OCR:**  ** OpenAI API (GPT-4o), Veryfi SDK
* **Data Orchestration:** `instructor` (Pydantic-based extraction), `pandas`
* **Preprocessing:** `pdf2image`, `Poppler`
* **Storage and Analysis:** Microsoft Excel (`openpyxl`)

---

## Python Scripts Explanation

<video
  controls
  preload="metadata"
  poster="https://quirkyagents.com/wizard/projects/pdf-excel-ai/python-script-ss.png"
  style="width:100%; height: auto; aspect-ratio: 16 / 9; border-radius:8px;"
>
  <source
    src="https://pub-21da2bddd9cf4685b4374f3e9115db36.r2.dev/python-script-pdf-ai.mp4"
    type="video/mp4"
  />
  Your browser does not support the vsideo tag.
</video>

###  OpenAI GPT-4o Integration
The OpenAI script utilizes the **`instructor`** library to enforce a strict **Pydantic schema**. Because GPT-4o is a reasoning model, 
the script uses specialized "System Prompts" to **handle mathematical extraction.**
* **Process:** Converts PDF to JPEG -> Base64 Encoding -> Vision-to-JSON extraction.
* **The Logic:** It doesn't just "read" text; it interprets it. When it sees "Item A 39.9x2", it calculates the unit price and quantity 
even if the receipt layout is cluttered.

###  Veryfi API Integration
The Veryfi script uses a dedicated financial OCR engine designed specifically for receipts and invoices.
* **Process:** Direct PDF/Image upload via the `veryfi` Python SDK.
* **The Logic:** It uses deterministic OCR rules to identify tax, subtotal, and vendor information instantly. It is pre-trained on 
millions of financial documents, making it **extremely fast** at identifying standard fields without complex prompting.

---

## Key Learnings & Conclusion

![Excel Output](https://quirkyagents.com/wizard/projects/pdf-excel-ai/excel-ai-output.png)

* **Reasoning vs. Recognition:** GPT-4o showed superior "reasoning" for non-standard multipliers, whereas Veryfi showed superior 
"recognition" for standard tax and total fields.
* **Accuracy vs. Cost:**
    * **OpenAI GPT-4o** is the more **economical** choice with slightly lower raw OCR precision but higher adaptability.
    * **Veryfi** is highly **accurate** and purpose-built for finance, but comes at a significantly **higher operational cost** per 
    document.

<br />
* ****The Verdict:**** GPT-4o is the optimal solution for cost-sensitive, high-volume projects, while Veryfi remains the benchmark for mission-critical financial auditing.

---