Navigation Bar

Logo
AnyParser API (YC S23) - The first LLM for document parsing with accuracy and speed | Product Hunt

Doubling Accuracy in Knowledge Retrieval from Charts and Tables

August 29, 2024
Back to Blogs
Authors: 
Jojo @  CambioML
AnyParser and Epsilla evaluation metrics from Ragas

Evaluation Metrics from Ragas

In today's data-driven landscape, industries like financial services rely heavily on precise and efficient information extraction from documents, especially those containing both unstructured text and structured data like tables and charts. Traditional Optical Character Recognition (OCR) models, despite their widespread use, often fall short in handling complex document formats, leading to suboptimal performance in advanced AI applications. Recognizing this gap, CambioML and Epsilla have introduced a cutting-edge knowledge retrieval system that promises to significantly enhance accuracy and recall in data extraction tasks.

Introduction: Overcoming OCR Limitations

OCR-based models, while effective at detecting text, struggle with extracting layout information and accurately pulling data from tables and charts. These limitations become particularly apparent in industries where precision is paramount, such as finance and healthcare. To address these challenges, CambioML and Epsilla have developed a novel approach that integrates state-of-the-art table extraction models with Retrieval-Augmented Generation (RAG) techniques. This new system achieves up to 2x precision and 2.5x recall compared to conventional RAG systems, setting a new standard for document question answering.

AnyParser: Revolutionizing Table Extraction

At the heart of this breakthrough is AnyParser, a model powered by advanced vision language models (VLMs) that excels in extracting information from diverse data sources. Unlike traditional models that rely heavily on OCR, AnyParser uses a combination of visual and text-based encoders to capture even the smallest details from documents, ensuring that no critical data is missed. This approach is particularly beneficial in extracting high-resolution data from financial and medical documents, where accuracy is critical.

Epsilla: A Flexible RAG Platform

Complementing AnyParser is Epsilla, a no-code RAG-as-a-Service platform designed to optimize various RAG pipelines. Epsilla enhances the knowledge retrieval process through advanced chunking, indexing, and query refinement techniques. By integrating keyword-based and semantic search methods, Epsilla delivers highly accurate and contextually relevant results, making it an ideal solution for large language model (LLM) applications.

Experiment & Evaluation: Real-World Impact

AnyParser and Epsilla evaluation metrics from Ragas

Evaluation Metrics from Ragas

To validate the effectiveness of AnyParser and Epsilla, the system was tested on 10-K financial documents from companies like Apple and Meta. The results were impressive, with the system demonstrating significantly higher performance across all key evaluation metrics, including context precision, recall, faithfulness, and answer correctness. In some cases, the system outperformed traditional RAG systems by as much as 2.7x, highlighting its superiority in handling complex data extraction tasks.

Common Use Cases and Key Benefits

  • Accuracy: High precision in converting both structured and unstructured data into usable formats.
  • Privacy: The ability to deploy the system within a customer’s data center ensures full data security.
  • Scalability: Rapid processing of large volumes of documents, enabling faster decision-making

Conclusion: A New Era in Knowledge Retrieval

The introduction of AnyParser and Epsilla marks a significant advancement in knowledge retrieval technology. By combining advanced extraction models with a robust RAG infrastructure, this integrated solution not only improves accuracy and efficiency but also offers the flexibility and privacy that modern enterprises demand. As technology continues to evolve, the applications and benefits of this system are vast and promising, making it a game-changer for industries that depend on precise data extraction.

For the full detailed whitepaper, please check out this link.


Footer