NovaLAD: A Fast, CPU-Optimized Document Extraction Pipeline for Generative AI and Data Intelligence

TL;DR

NovaLAD is introduced, a comprehensive document parsing system that integrates two concurrent YOLO object detection models - element detection and layout detection - with rule-based grouping and optional vision-language enhancement.

Abstract

Document extraction is an important step before retrieval-augmented generation (RAG), knowledge bases, and downstream generative AI can work. It turns unstructured documents like PDFs and scans into structured text and layout-aware representations. We introduce NovaLAD, a comprehensive document parsing system that integrates two concurrent YOLO object detection models - element detection and layout detection - with rule-based grouping and optional vision-language enhancement. When a page image is sent in, the first thing that happens is that it goes through both models at the same time. The element model finds semantic content like the title, header, text, table, image, and so on, and the layout model finds structural regions like layout_box, column_group, multi_column, row_group, and so on. A key design decision is to first send an image or figure through an image classifier (ViT) that decides whether it is relevant or not. Only useful images are then submitted to the Vision LLM for title, summary, and structured information, which cuts down on noise and costs. NovaLAD is built for speed: it works on CPU, employs parallel execution for detection, classification, OCR, and conversion, and generates several forms, including structured JSON, Markdown, RAG-ready texts, and knowledge graphs. We test on the DP-Bench benchmark (upstage/dp-bench) and get 96.49% TEDS and 98.51% NID, which is better than both commercial and open-source parsers. This paper explains how to extract data, how the architecture works, how data flows, and how to make NovaLAD both accurate and usable without needing a GPU.

Figures (8)

• Figure 1: Complete end-to-end NovaLAD pipeline. The process begins with a document page fed to parallel Layout Detection and Elements Detection (YOLO models). Images/figures are filtered by a useful vs useless classifier. Layout and filtered elements merge into a unified representation, followed by OCR content extraction. If an LLM is configured, advanced extraction of titles, summaries, and data is performed. The pipeline outputs structured JSON, which is converted to multiple formats (document chunks, Markdown, knowledge graph) with document category classification, culminating in the final NovaLAD output.
• Figure 2: Training and validation dynamics for the layout detection YOLO model. (a) shows how precision, recall, mAP50, and mAP50--95 improve over 30 epochs; (b)--(e) decompose the box, classification, DFL, and objectness losses on train and validation splits; (f) visualizes the learning-rate schedule used during training.
• Figure 3: Training and validation dynamics for the element detection YOLO model. (a) shows precision, recall, mAP50, and mAP50--95 over 50 epochs; (b)--(e) detail how box, classification, DFL, and objectness losses evolve on train and validation sets; (f) shows the corresponding learning-rate schedule.
• Figure 4: Training and validation behaviour of the ViT image classifier. Panels (a)--(c) show that loss decreases monotonically while accuracy rises to 98.53%; (d) tracks the gap between training and validation performance, which remains small throughout; (e)--(f) highlight that convergence is fast (4 epochs) and that the final operating point lies in the regime of low validation loss and high accuracy.
• Figure 5: Novalad Output post Element and Layout Detection