Wedha Strengthens Generative AI Quality Control With RAGOps-Based ‘VectorGreen’

Wedha announced on the 30th that it has developed in-house ‘VectorGreen,’ a RAGOps-based LLM quality management solution. As enterprise use of generative AI expands into real-world business areas such as finance and manufacturing, the company said the importance of a system that manages not only model performance but also data freshness, retrieval accuracy, and the basis for answer generation is growing.

‘VectorGreen’ supports the accuracy, freshness, and provenance tracking of answers generated by enterprise generative AI. The company said the move reflects the view that in customer service and financial work environments, incorrect answers directly affect service quality, making it necessary to check the entire operating process.

‘VectorGreen’ was designed based on a RAGOps structure that manages the operation of retrieval-augmented generation (RAG) systems. Its management scope covers the entire process of data collection, retrieval, answer generation, quality evaluation, and history tracking.

The function supports the automatic collection and reflection of changed data, and continuously incorporates data updates to help generative AI provide answers based on the latest information. Through this, it aims to reduce errors caused by outdated information and inaccurate retrieval.

It also moves beyond simply checking whether an answer was generated through its LLM quality evaluation function to inspect the retrieval and generation processes at the same time. It evaluates whether data appropriate to the question was retrieved, whether the retrieved information was properly used in answer generation, and whether the final answer matches the reference data.

With this structure, hallucination, incorrect retrieval results, and inaccurate answers can be continuously checked. In addition, it is characterized by integrated management of a company’s own data, retrieval systems, prompts, and answer results.

Wedha provides a Query-Answer Lineage function. This function allows users to trace and verify the path from a user question to retrieval data, reference evidence, and the final answer generation process. Users can see not only the result, but also what data and evidence the answer was based on.

This function allows financial clients to track source data, retrieval results, queries, and response histories in addition to the final answer when asking about a specific product or business procedure. If an error occurs, it also makes it possible to identify the problematic stage. It can also be used to revise related data and answers.

Wedha will first apply this function to the financial sector with VectorGreen. It plans to expand its application to manufacturing, the public sector, and other areas afterward. It also intends to advance the system toward integrating generative AI data management, quality evaluation, operations, and monitoring functions into a single RAGOps platform.

Shin Dong-min, head of Wedha's Technology Research Institute, said that as enterprise adoption of generative AI expands, it is important to continuously manage not only AI model performance but also the freshness of company data and answer quality. He added that the company plans to support trusted operation of enterprise generative AI through automatic collection of changed data, as well as RAGOps-based LLM quality evaluation and Query-Answer Lineage.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407511

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407511