A Hybrid Pipeline using PaddleOCR Layout Analysis & LLM Text-to-SQL

This talk demonstrates a pipeline using PaddleOCR for PDF layout analysis to structure financial data, followed by an LLM generating SQL queries against a database.

Overview

I will demonstrate a live, end-to-end pipeline that solves the nightmare of unstructured financial data. The demo consists of two technical deep-dives:

The “Jailbreak” (PDF => Structured): I’ll show how to use PaddleOCR-VL not just for text detection, but for Layout Analysis. We will take a messy, multi-column financial PDF, identify the regions (headers vs. paragraphs vs. tables), and reconstruct the tables directly into Pandas DataFrames while preserving row/column integrity—something standard OCR fails at.

The “Interrogation” (Natural Language => SQL): Once the data is structured (simulated in MySQL via Docker), I’ll demo a Python agent using ERNIE 5.0 (via an OpenAI-compatible client) that acts as a “SQL Architect.” It takes natural language questions from a CFO persona (e.g., “How did T-103 perform against targets in Asia?”), generates valid SQL, executes it against the containerized DB, and interprets the results.

Tech stack