Automatically structure and add metadata to large amounts of unstructured documents, including tables and figures.
Enabling highly accurate answer generation
~ Trial version available from November 6th, enabling low-cost testing similar to production ~

　
We have developed a proprietary multi-modal generative AI product called "Super RAGWe are planning to release a packaged version of our latest AI application, Super Lag™ (Super Lag), which will come standard with a variety of AI functions, including document graph construction, in mid-January 2025. In addition, we will offer a trial version of the product from November 6th, which will enable low-cost verification similar to that of a production version.

Since March of this year, we have been providing "Super RAG," which can generate answers and documents by automatically structuring large amounts of complex, unstructured documents (sales brochures with charts and diagrams, administrative procedures, internal regulations, etc.) that are difficult to handle with general RAG, customized for each company. "Super RAG" combines three unique technologies, "document structure analysis," "unique search system," and "prompt optimization," with existing LLM services such as Azure OpenAI Service to generate accurate and advanced answers. By incorporating generative AI that can utilize a huge amount of in-house data into various business scenes, we will promote essential DX and hyper automation, such as extracting information from complex documents and automatically generating materials.

The newly announced "Super RAG" package version not only automatically structures and imports complex unstructured documents, but also provides a one-stop service that enables the creation of AI assistants without coding, high-precision searches using document graphs, and connections to various systems and applications.

[Notice] About the trial version of "Super RAG"

In preparation for the sale of the packaged version of "Super RAG" in mid-January 2025, we will begin offering a trial version for performance verification and trial operation from November 6th. The trial version allows companies to utilize their existing application functions while also providing a mechanism for using "Super RAG's" various AI systems at low cost and in a short period of time (approximately one month). Please take advantage of this opportunity to check out the technological advantages of "Super RAG" through trial operation.

▼"Super RAG" materials download Click here
▼For corporate inquiries regarding "Super RAG" Click here

■ "Super RAG" enables low-cost, versatile LLM to be introduced into business

RAG (Retrieval Augmented Generation) is attracting attention as a way to promote the introduction of large-scale language models (LLMs) such as ChatGPT in actual business operations. RAG is a technology that improves the accuracy of answers by appropriately combining external information when LLMs generate text. The proprietary information held by companies contains a lot of documents in complex formats (unstructured documents) such as tables and graphs, and advanced AI technology is required to input this data and achieve the accuracy required for business use.

Currently, trial implementation of systems using RAG is underway, primarily at large companies, but to improve response accuracy, individual customization is required for each use case, such as data preprocessing and search algorithms. As a result, implementation is limited to only some operations from the perspective of response accuracy and ROI, and has not yet been implemented company-wide, including for low-volume, high-mix operations (long-tail operations), and the current situation is that internal usage rates remain low. In addition, there are issues such as preventing the implementation of a system that is heavily dependent on the technology of a specific vendor and the risks involved in transferring large amounts of data, making it difficult to completely replace the in-house RAG system currently in use.

Cinnamon AI's "Super RAG" addresses these issues by using its unique AI technology (a group of AI modules) that can be securely and easily integrated into existing in-house systems to generate generic, highly accurate answers at low cost. For tasks where the accuracy was around 401 TP3T when using a general RAG, "Super RAG" achieved an accuracy of over 901 TP3T (based on our test data).

▽ "Super RAG" - solving the role and issues of processes in RAG

▽ "Super RAG" technology examples

Conventional RAG technology performs vector searches using semantic similarities between texts, but in many cases it ends up searching for unrelated content with similar character strings, which is the cause of reduced accuracy of answers (see left in figure below). "Super RAG" analyzes the document structure, then divides and classifies it into content units, and then structures the relationships between each piece of content in graph form and stores it in a database, thereby improving accuracy by performing searches using the relationships between content (see right in figure below).

▽ Examples of documents that "Super RAG" is good at analyzing

▽ Examples of use cases for "Super RAG" (including customization)

▼Contact information for companies regarding this matter
https://contents.cinnamon.ai/contact/inquiry

*"Super RAG™" mentioned in this press release is a trademark of Cinnamon Co., Ltd.