AI Innovations and Insights 27: OCR Hinders RAG and RAGChecker

Author(s): Florian June

Originally published on Towards AI.

This member-only story is on us. Upgrade to access all of Medium.

This article is the 27th in this mind-expanding series.

Today, we will explore two enlightening topics in AI, which are:

OCR Hinders RAG: Unraveling the Distorted Knowledge PuzzleRAGChecker: A Careful Teacher

Open-source code: https://github.com/opendatalab/OHR-Bench

Imagine RAG as a system trying to piece together a complete world jigsaw puzzle, but the pieces it receives from OCR are missing, distorted, or sometimes even belong to an entirely different puzzle.

For example, semantic noise changes the color and shape of certain pieces — like turning “E=mc²” into “E=mc³” — while formatting noise alters their structure, cutting rounded pieces into squares, as seen when table formats get scrambled.

As a result, RAG assembles a distorted version of the world, leading to inaccurate or even absurd answers.

“OCR Hinders RAG” examines how these “missing pieces” affect knowledge extraction and proposes better strategies to complete the puzzle.

We know that PDF parsing is a key part of RAG.

However, OCR introduces noise when extracting information from unstructured PDFs, and since RAG systems are sensitive to input quality, OCR errors can have a cascading impact on knowledge base construction and retrieval.

Figure 1. Construction of OHRBench and evaluation protocol. [Source].

As shown in Figure 1, “OCR Hinders… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI