Kingsoft Office Zhu Yie: From "Seeing" to "Mastering": AI Applications Enter the Era of "Knowledge-Enhanced Generation" | Alpha Summit

On December 20th, at the "Alpha Summit" jointly organized by Wallstreetcn and China Europe International Business School, Zhu Yie, Assistant Vice President and Senior Technical Expert of Kingsoft Office, delivered a speech titled "WPS AI: Moving Towards Higher-Quality Knowledge-Augmented Generation."

He stated that the core challenge of current AI applications has shifted from competition in model capabilities to how to efficiently utilize enterprise private data. The convergence of model capabilities means it is difficult to form a monopoly advantage.

He emphasized that the key to truly determining the value of AI applications lies in transforming a large quantity of complex, unstructured document data within enterprises into high-quality, model-understandable knowledge assets. Traditional RAG faces fundamental limitations such as "documents are not equivalent to knowledge" and "semantic similarity is not equal to logical relevance," so it is necessary to promote an upgrade in the technical paradigm from "model-centric" to "data/knowledge-centric."

He stressed that the future path is to develop "KAG (Knowledge Augmented Generation)." This requires enterprises to systematically manage, model, and apply knowledge as they manage data. Specifically, it's necessary to integrate multi-modal, multi-structured knowledge through technologies like VLM and knowledge graphs, and to build a dual architecture of "data lake" and "knowledge lake." The ultimate goal is for AI to truly "master"—not just "see"—enterprise knowledge, so as to deliver reliable value in scenarios such as professional Q&A, intelligent writing, and compliant creation, thus achieving a key leap from digitalization to intelligence.

The following are highlights from his speech:

Enterprise AI applications are shifting from “model-centric” to “data-centric.” Data quality has become the key factor determining the effectiveness of AI applications. WPS AI aims for Knowledge-Augmented Generation, helping large models truly “master” a company’s knowledge assets.

Manage knowledge as you manage data. Transforming data and knowledge into AI-usable assets is the foundation for enterprises to move from digitalization to intelligence. In the DATA 2.0 era, enterprises need to manage knowledge as they manage data. WPS 365 builds an exclusive “enterprise brain” for clients through knowledge modeling, knowledge governance, and multimodal fusion.

High-quality output begins with high-quality input. If the input is chaotic or conflicting raw data, no matter how powerful the model, output cannot be reliable. Therefore, knowledge governance is the cornerstone for the implementation of AI in professional domains, whose importance will surpass algorithm optimization itself.

Specialized AI applications are a "knowledge engineering" process, not a simple technical integration. From writing compliance reports to precise information extraction, their essence is the systematization and structuring of professional domain knowledge. Whoever upgrades their knowledge assets first will gain a true competitive edge in the AI era.

True intelligence is not just “seeing” documents but “understanding” logic. Mainstream AI applications (such as RAG) encounter bottlenecks because “semantic similarity is not equal to logical relevance.” The real breakthrough lies in integrating knowledge graphs and business rules from multiple sources, enabling logical reasoning and precise answers to unlock value in professional scenarios.

The following is a summary of the key content by Wallstreetcn:

After Large Models, What Is the Real Bottleneck?

A major consensus currently is that the general intelligence of advanced large models in knowledge reserve and logical understanding has surpassed ordinary employees, and model capabilities have converged, making monopoly hard. So, the core question shifts to: How can large models truly realize their value in practical applications?

Our answer: Deep integration with external data, especially private enterprise data, is essential. However, data in the form of "documents" is not equivalent to "knowledge," because a vast enterprise document pool (texts, spreadsheets, PDFs, etc.) suffers from complex formats, disorganized structures, and contradictory content. For instance, one document may state an unused annual leave conversion rate of 200%, while another says 300%; one regulation may require six months of data retention, while another says only necessary data must be retained. If these conflicts aren't resolved, AI output is unreliable.

The deeper challenge lies in the mainstream technical paradigm. The widely used RAG (Retrieval-Augmented Generation) relies on “vector similarity retrieval.” This has a fundamental limitation: Semantic similarity does not equal logical relevance. For example, when asked "What to do if the laptop won't turn on," the system may retrieve a document detailing "MacBook Pro 14-inch" specifications (semantically similar), but miss a troubleshooting guide that never mentions the word "laptop" yet truly solves the problem (logically relevant). This results in many AI applications looking great in demos but struggling in production.

From RAG to KAG—Building a New Paradigm for Knowledge-Augmented Generation

To break these bottlenecks, we propose evolving from RAG to KAG (Knowledge Augmented Generation). This is not merely an optimization but a paradigm shift, with two core points:

First, high-quality input leads to high-quality output. Knowledge must be governed to resolve conflicts, fill gaps, and build structure.

Second, it's necessary to systematically integrate multi-modal and multi-structured knowledge assets. Retrieval cannot be limited to documents; it must also integrate existing knowledge graphs, structured tags, workflow SOPs, etc., to provide high-quality input for AI generation.

Based on this, we designed a two-layer architecture. The bottom is the "knowledge governance layer," responsible for document parsing, knowledge extraction, graph building, and quality monitoring; the top is the "knowledge application layer," integrating multi-source retrieval engines, dynamic sorting modules, and context engineering, to build a knowledge base empowering professional scenarios.

Implementing KAG in Four Key Scenarios

Based on the KAG architecture, we developed an intelligent document repository product, focusing on four core scenarios:

First, knowledge governance. Automated knowledge extraction and graph building help clients detect duplicate content, logical conflicts, and missing knowledge in their document repositories. For example, the system can automatically flag two conflicting versions for annual leave conversion percentage, or note missing key sections like "printer driver installation" in the "IT support" knowledge base, assisting administrators in decision-making and optimization.

Second, professional intelligent Q&A. After integrating private domain document graphs with industry regulations, SOPs, and other structured knowledge, our Q&A system can handle complex professional queries. For example, a user may ask: "In Zhejiang Province, is it permissible to use ingredient X when manufacturing a certain particle size of pharmaceutical raw material? Please reference only regulations for the year 2025." The system can accurately parse constraints such as location, ingredient, and year, and deliver precise answers.

Third, intelligent extraction from complex documents. We optimized for complex tables, checkboxes, and handwriting often seen in medical reports, contracts, and invoices. A pharmaceutical client uses this feature to automatically parse and extract key fields from drug adverse event report email attachments, populating their medication management system. Tasks previously taking hours are now reduced to minutes.

Fourth, intelligent writing in professional fields. This is not writing a simple leave request, but drafting industry reports (e.g., clinical study report CSR) with strict format and data citation requirements. We use two collaborative agents: One generates an "intelligent template" covering outline and data needs based on samples and regulations; the other references the template to accurately and losslessly extract required data and tables from massive experimental datasets, finally generating a format-compliant, data-accurate professional report—shortening writing cycles from weeks to days.

Manage Knowledge as You Manage Data

Finally, I would like to conclude. The evolution from RAG to Graph RAG to KAG is an upgrade from “letting large models see documents” to “understanding the logic between documents” to “truly mastering enterprise knowledge assets.”

We believe that in the age of intelligence, enterprises need to build a dual architecture of “data lake” and “knowledge lake.” In the future, enterprises must not only accumulate raw data, but also systematically conduct knowledge management, knowledge modeling, and knowledge governance, just as they manage data today. This will be the key foundation for enterprises transitioning from digitalization to intelligence, and it is the only way for AI to truly deliver efficiency improvements in professional domains.

Risk Warning and DisclaimerThe market involves risks, and investments must be made cautiously. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in the article suit their specific circumstances. Investments based on this content are at your own risk.