Promote DX by importing internal documents as they are.
Automatically reads complex layouts, tables, charts, bar graphs, diagrams, and handwritten notes often found in Japanese documents..

　
Our company announced Super RAG, a document LLM that significantly improves business efficiency and leverages knowledge by utilizing internal documents. By utilizing "Super RAG", it is possible to extract information from complex forms including complex tables, charts, bar graphs, diagrams, and handwritten documents while suppressing hallucinations and supporting the use of all kinds of documents when using LLM. To do.

Super RAG: https://cinnamon.ai/ai_model/super-rag/
Super RAG material DL form: https://go.cinnamon.ai/WP_SuperRAG_DL_LP.html

*Details of output examples are listed below.

■ Background and overview of document LLM “Super RAG” development

Retrieval Augmented Generation (RAG) is attracting attention as a method to promote the introduction of large-scale language models (LLM), which are increasingly used in companies, into actual business operations. RAG is a technology that improves answer accuracy by combining text generation using LLM with external information retrieval, and can be translated as ``search extension generation'' or ``retrieval extension generation.'' By combining external information searches, the effect of having LLM generate answers based on the latest accurate information and the basis of output results becomes clear, suppressing the phenomenon of generating information that is not based on facts (hallucination) It is expected that this will have the following effects.

Since the introduction of RAG, many companies have tried to expand the possibilities of LLM by utilizing RAG. Many challenges have been recognized in the process of importing data into RAG from the perspective of document analysis, understanding, and meaning extraction, such as the need for knowledge and the difficulty of inputting prompts. This is hindering progress. In addition, RAG itself is provided as an application by many companies, making technical verification easy, but the reality is that no solution has emerged that fundamentally solves the issues mentioned above.

Until now, Cinnamon AI has accumulated technical assets that handle unstructured data, such as AI-OCR, natural language processing, and speech recognition, and has applied them to intelligent document processing (analyzing documents and extracting and organizing information). By integrating advanced technology (hereinafter referred to as IDP), we have provided advanced document analysis and meaning extraction technology. In addition to these IDP technologies, by incorporating proprietary technologies such as the utilization of domain knowledge and automatic prompt generation, we have developed "Super RAG," Cinnamon AI's unique RAG system. With three unique technologies: document analysis, knowledge injection, and automatic prompt generation, we have overcome the challenges of RAG (document complexity, lack of domain knowledge, and reduced accuracy due to ambiguous prompts) and greatly facilitated the use of LLM by companies. Masu.

■ Features of Document LLM “Super RAG”

Document LLM "Super RAG" provided by Cinnamon AI is based on the fact that all the functions provided by LLM can be used, and by utilizing three unique technologies such as document analysis, knowledge injection, and automatic prompt generation, We are able to import all company documents. In particular, we have realized the extraction of information from complex forms, including complex tables, charts, bar graphs, diagrams, and handwritten documents, which were previously considered difficult to analyze.

Below, we will introduce samples of charts and tables that are difficult to read using the general RAG system, as well as information extraction results using the Cinnamon AI "Super RAG". The original data to be read is sample data created by our company, but the reading results are the output results of actually inputting each data into "Super RAG".

Example of information extraction using “Super RAG”

■ Promote collaboration with end-user companies and partner companies

The document LLM "Super RAG" will be sold to end-user companies that are considering using LLM in-house, as well as partner companies that support end-users' use of LLM through existing applications and customization.

Cinnamon AI has provided many AI solutions to major domestic companies. Generative AI makes it easy to use solutions provided by various companies in an interactive chat format, and end-user companies have begun using it on their own tests. On the other hand, it is recognized that there are many challenges to practical application, such as controlling hallucination and utilizing in-house documents and knowledge.

Partner companies are looking into partnerships in which "Super RAG" is incorporated into provided applications and jointly provided, and partnerships in which "Super RAG" is adopted when supporting end users' use of LLM through customization. We have started collaboration with.

In the future, Cinnamon AI will actively promote the Document LLM "Super RAG" for end user companies and partner companies.