Skip to main content

Introduction

BETA Research Work in Progress

The DeepSeek OCR project evaluates the performance of DeepSeek-VL (Vision Language) models for Optical Character Recognition, particularly focusing on complex document structures and Sri Lankan legal texts.

Documentation​

Findings​

  • Experiments - DeepSeek OCR vs Gemini 2.0 comparison study

Technology​

  • Setup - Installation and environment configuration
  • Usage - How to use the OCR library

Objectives​

  • Benchmark Accuracy: Evaluating text extraction quality against standard OCR tools.
  • Structured Extraction: Successfully demonstrating the ability to extract organizational charts and tabular data directly into JSON.
  • Optimization: Re-using the loaded model for both OCR and post-processing to minimize VRAM usage.

Architecture​

The project has evolved into a reusable Python library ldf.deepseek.ocr that supports:

  1. Batch Processing: Efficiently processing folders of PDF documents.
  2. Multi-Agent Workflow: A sophisticated pipeline involving:
    • Extractor: Runs the core DeepSeek-OCR model to get raw text.
    • Processor: Structures raw text into JSON based on custom prompts/schemas.
    • Aggregator: Stitches multi-page documents together, handling context continuation.
    • Finalizer: Validates and saves the consolidated output.