November 2023 Summaries

Independent Health collaborated with LangChain and Unstructured to address the complexities of managing the "Certificate of Coverage" (CoC), a detailed document crucial for policyholders. By exploring a Retrieval-Augmented Generation (RAG)-based architecture, they aimed to simplify the process of answering insurance policy questions. The study compared a baseline RAG with a nuanced Semi-Structured RAG architecture that utilizes Unstructured's data ingestion and transformation capabilities along with LangChain's Multi-Vector Retrieval. This approach effectively processes semi-structured data from CoC documents, which often consist of structured tables mixed with natural language text. The evaluation, conducted using LangChain’s platform LangSmith, demonstrated that the Semi-Structured RAG outperformed the baseline by accurately answering three out of four questions, highlighting the importance of tailored storage and retrieval strategies for semi-structured data. This initiative underscores the potential of advanced data processing tools to enhance document handling in healthcare insurance, ultimately improving service quality for policyholders.

Unstructured's content-aware chunking method enhances the performance of Retrieval-Augmented Generation (RAG) applications by producing more coherent and contextually relevant document segments than traditional character-based chunking. This approach improves the quality of LLM outputs by ensuring that chunks have a consistent semantic meaning, which is crucial when dealing with content spread across multiple sections or documents. In a test involving 68 documents, outputs generated using Unstructured's chunking were deemed more relevant than those from standard chunking two-thirds of the time. This method not only results in more precise and detailed responses but also allows for more accurate citations, as demonstrated in a comparison of responses to a query about the Fresno-Merced Future of Food coalition. The Unstructured chunking facilitated a more comprehensive and specific answer, illustrating its effectiveness in producing higher fidelity RAG outputs.

Unstructured introduces a content-aware chunking strategy for Retrieval-Augmented Generation (RAG) systems, offering higher quality outputs compared to traditional character-based chunking. This approach identifies document elements like titles and body text to create coherent segments, leading to more relevant responses and precise citations in natural language applications. In a test using GPT-4, Unstructured chunking proved more effective in providing detailed and comprehensive responses, especially when content is dispersed across multiple sections or documents. This method enhances the retrieval process by focusing on semantically consistent chunks, improving both the relevance of query results and the number of citations in responses.

As of 2023, Unstructured has significantly expanded its offerings to support engineers in connecting diverse datasets to large language models (LLMs), following the widespread adoption of ChatGPT. The company is focusing on three product lines: the open-source library, a paid production API, and the upcoming Unstructured Enterprise Platform. The open-source library remains an accessible option for developers creating prototype applications, while the production API offers enhanced features such as advanced PDF and image processing and is available through the Azure marketplace, with AWS availability forthcoming. The anticipated Enterprise Platform aims to provide a comprehensive ETL solution with features like job scheduling and advanced data chunking, designed to facilitate the efficient transformation of LLM applications into business tools. Unstructured is committed to refining its tools for enterprise-level reliability and encourages engagement through social media and community platforms to gather user feedback and insights.

As 2023 unfolds, Unstructured has capitalized on the rising interest in large language models (LLMs) by offering tools that facilitate the connection of diverse datasets to these models. With the introduction of a free API, the company's open-source package has seen a surge in usage, supporting thousands of users and organizations in managing complex data processing tasks. Unstructured aims to refine its tools for enterprise-grade reliability as businesses begin to operationalize LLM applications, focusing on three product lines: open source, API, and an upcoming enterprise platform. The open-source library remains a reliable resource for developers building prototype applications, while the paid API offers enhanced performance for production-ready tools, featuring advanced data processing capabilities and secure data handling on platforms like Azure. The forthcoming Unstructured Enterprise Platform promises a comprehensive ETL experience with features such as job scheduling, monitoring, and automated data ingestion, designed to optimize the performance of downstream retrieval-augmented generation (RAG) systems. The company encourages engagement through social media and community channels to stay updated and provide feedback on their evolving tools.