On December 20th, at the "Alpha Summit" jointly hosted by Wall Street News and CEIBS, Zhu Yi'e, Assistant President and Senior Technical Expert of Beijing Kingsoft Office Software,Inc., delivered a keynote speech titled "WPS AI: Towards Higher-Quality Knowledge-Augmented Generation."

He stated that the core challenge for current AI applications has shifted from competition in model capabilities to how to efficiently utilize enterprise private data. The convergence in model capabilities implies that they are inherently difficult to monopolize as a competitive advantage. He emphasized that the true key to determining the value of an AI application lies in transforming the vast amount of complex, unstructured document data within an enterprise into high-quality knowledge assets that can be understood by models. Traditional RAG faces fundamental limitations—"a document is not equivalent to knowledge" and "semantic similarity does not equal logical relevance"—necessitating a technological paradigm shift from being "model-centric" to becoming "data/knowledge-centric."

He stressed that the future path involves developing "KAG (Knowledge-Augmented Generation)." This requires enterprises to systematically govern, model, and apply knowledge, much like they manage data. Specifically, it necessitates leveraging technologies like VLM and knowledge graphs to integrate multimodal, multi-structured knowledge, and to build an architecture that equally prioritizes both a "data lake" and a "knowledge lake." The ultimate goal is to enable AI to truly "master" enterprise knowledge, not just "see" it, thereby delivering reliable value in scenarios such as professional Q&A, intelligent writing, and compliant content creation, and completing the critical leap from digitalization to intelligence.

Key insights from the speech are as follows:

Enterprise AI applications are transitioning from being "model-centric" to "data-centric." Data quality has become the key determinant of AI application effectiveness. WPS AI aims for Knowledge-Augmented Generation to help large models truly "master" an enterprise's knowledge assets.

Manage knowledge like you manage data. Transforming data and knowledge into AI-usable assets is the cornerstone for enterprises moving from digitalization to intelligence. In the DATA 2.0 era, enterprises must manage knowledge as diligently as they manage data. Through knowledge modeling, knowledge governance, and multimodal integration, WPS 365 helps build a proprietary "enterprise brain" for companies.

High-quality output must begin with high-quality input. If the input consists of chaotic, conflicting raw data, the output will be unreliable regardless of the model's power. Therefore, knowledge governance is the foundation for AI implementation in professional fields, and its importance will surpass algorithm optimization itself.

The professional application of AI is a "knowledge engineering" endeavor, not merely a technical integration. From drafting compliance reports to precise information extraction, the essence is the process of systematizing and structuring knowledge from professional domains. Whoever率先completes the upgrade of their own knowledge assets will be able to establish a real competitive advantage in the AI era.

True intelligence is not about "seeing" a document, but "understanding" the logic. Current mainstream AI applications (like RAG) face bottlenecks because "semantic similarity does not equal logical relevance." The real breakthrough lies in integrating multi-source knowledge, such as knowledge graphs and business rules, enabling AI to perform logical reasoning and provide precise answers, thereby unlocking value in professional scenarios.

Following is the essence of the content:

After large models, what is the real bottleneck? A key consensus is that the comprehensive intelligence of cutting-edge large models has surpassed that of the average employee in terms of knowledge reserves and logical understanding, and model capabilities are converging, making monopolies difficult. The core question then becomes: How can large models deliver real value in practical applications?

The answer is that they must deeply integrate with external data, especially enterprise private data. However, data existing in "document" form is not equivalent to "knowledge" due to inherent shortcomings in the vast amount of enterprise documents—complex formats, disorganized structure, and contradictory content. For instance, one document might state an untaken annual leave conversion rate of 200%, while another says 300%; one regulation requires storing data for six months, another says to retain only necessary data. If these conflicts are not resolved, AI output will be unreliable.

A more profound challenge lies in the mainstream technological paradigm. The core of the widely used RAG (Retrieval-Augmented Generation) technology is "vector similarity retrieval." This introduces a fundamental limitation: semantic similarity does not equal logical relevance. For example, asking "What should I do if my laptop won't turn on?" might retrieve a document detailing the specifications of a "MacBook Pro 14-inch" (semantically similar) but miss a genuine troubleshooting guide that doesn't mention the word "laptop" (logically relevant). This leads to many AI applications being "impressive in demo but difficult in production."

From RAG to KAG: Building a New Paradigm of "Knowledge-Augmented Generation" To break through these bottlenecks, it is necessary to evolve from RAG to KAG (Knowledge-Augmented Generation). This is not a simple optimization but a paradigm shift. Its core tenets are twofold:

First, high-quality output requires high-quality input. Knowledge must first be governed to resolve conflicts, fill gaps, and establish structure. Second, there must be systematic integration of multimodal, multi-structured knowledge assets. It's not enough to just retrieve documents; existing enterprise knowledge graphs, structured tags, process SOPs, etc., must be fused to provide high-quality input for AI generation.

Based on this, a two-tier architecture has been designed. The underlying layer is the "Knowledge Governance Layer," responsible for document parsing, knowledge extraction, graph construction, and quality monitoring. The upper layer is the "Knowledge Application Layer," which, with a multi-source fusion search engine, dynamic ranking module, and context engineering system as core components, builds a knowledge base that empowers various professional scenarios.

Implementing KAG in Four Key Scenarios Based on the KAG architecture, an intelligent document library product has been developed, focusing on four core scenarios:

First, Knowledge Governance. Through automated knowledge extraction and graph construction, the system helps clients identify duplicate content, logical conflicts, and knowledge gaps in their document libraries. For instance, it can automatically flag two conflicting versions regarding annual leave conversion rates or point out that an "IT support" knowledge base is missing a crucial section on "printer driver installation," assisting administrators in decision-making and optimization.

Second, Professional Intelligent Q&A. After integrating private document graphs with structured knowledge like industry regulations and SOPs, the Q&A system can handle complex professional queries. For example, a user can ask: "In Zhejiang Province, for producing an active pharmaceutical ingredient of a specific particle size, can component X be used? Please base your answer solely on 2025 regulations." The system can accurately parse multiple constraints like location, component, and year, and provide a precise answer.

Third, Intelligent Extraction from Complex Documents. Specialized optimizations have been made for complex tables, checkboxes, handwritten text, etc., commonly found in medical reports, contracts, and invoices. A pharmaceutical client used this feature to automatically parse email attachments of adverse drug reaction reports, extract key fields, and feed them back into the client's drug management system, reducing a task that previously took hours manually to just minutes.

Fourth, Intelligent Writing in Professional Domains. This differs from writing a leave request; it involves drafting industry reports with strict formats and precise data citation requirements, such as Clinical Study Reports (CSR). Two intelligent agents work collaboratively: one is responsible for generating a "smart template" containing an outline and data requirements based on examples and regulations; the other, following the template, precisely locates and losslessly cites the required data and tables from massive experimental data, ultimately generating a professionally formatted and accurate report, significantly shortening the drafting cycle from weeks.

Manage Knowledge Like You Manage Data In conclusion, the evolution from RAG to Graph RAG to KAG represents an upgrade from "enabling large models to see documents," to "understanding the logic between documents," and finally to "truly mastering enterprise knowledge assets."

It is believed that in the intelligent era, enterprises need to build a new architecture that equally emphasizes both a "data lake" and a "knowledge lake." In the future, enterprises must not only accumulate raw data but also systematically conduct knowledge operations, knowledge modeling, and knowledge governance, just as they have managed data in the past. This will be the key cornerstone for enterprises transitioning from digitalization to intelligence and the essential path for AI to truly enhance efficiency in professional fields.