[DigitalToday reporter Chi-gyu Hwang] Signs of change are emerging in retrieval-augmented generation (RAG), long considered a standard for building enterprise artificial intelligence (AI).

Google Cloud recently released Open Knowledge Format (OKF), proposing a way to manage core enterprise knowledge as a structured knowledge network rather than storing it in a vector database for retrieval.

According to a post recently shared on Medium by the tech blog Secret Dev, RAG splits enterprise PDFs and documents into small units, embeds them, then finds information close in meaning to a question and provides it to a large language model (LLM). It has the advantage of searching broadly across various documents, but tables and context can be damaged in the process of splitting documents, and outdated information may be retrieved. If data changes frequently, the burden of regenerating embeddings and managing indexes also grows.

As an alternative to these problems, OKF supports building enterprise knowledge like a single "living wiki". Its core unit, a "Knowledge Bundle", is a directory made up of YAML metadata and Markdown documents rather than a separate proprietary database. The directory path defines the location of a concept, and each document records information such as the type of knowledge, the person in charge, the time of revision and sources.

The biggest difference is how information is connected. While RAG probabilistically finds highly relevant documents by calculating similarity between vectors, OKF searches for needed knowledge by following links specified in documents. For example, after checking definitions of key indicators such as "sales" or "churn rate", it proceeds in order to linked database schemas and calculation rules.

This allows core business rules and data definitions to be managed more accurately, Secret Dev said.

If a developer changes the database structure, an AI agent can also be set up to update related documents and record the change history. Because all knowledge remains as plain text files, version control and audits through Git are also possible.

OKF does not completely replace RAG.

Secret Dev explained: "RAG remains useful for information that has a broad scope and whose exact location is difficult to define in advance, such as past reports, customer inquiries and vast unstructured materials. By contrast, for information such as tax identification numbers, sales definitions and database schemas, where mistakes directly cause business problems, OKF, which is managed explicitly, is suitable." It added: "It is highly likely that the knowledge management structure for enterprise AI will evolve toward combining the two technologies rather than choosing one of RAG and OKF. An AI router will judge the nature of a question, finding core rules and the latest schemas in OKF, while searching vast past materials or unstructured information in RAG."