Recently, the ZEHO ECO Knowledge Base completed testing and was officially launched. This achievement marks a critical step in the company's intelligent transformation within the ecological environment field. The core system of this knowledge base, the "AI Large Model Assistant," deeply integrates ZEHO ECO's 28 years of industry experience and the foundation of the Zhipu GLM large model. It combines the dual AI models of Zhipu Z1 and DeepSeek R1 to build an intelligent interactive hub covering the entire chain of tendering/bidding, planning & design, engineering construction, and internal control management. It enables information retrieval within seconds and intelligent, precise push notifications, providing strong support for efficient operations, intelligent decision-making, innovation capabilities, as well as market expansion and customer service.

ZEHO ECO AI Large Model Assistant

01 Building an Industry-Leading Knowledge Base with a Clear Knowledge Architecture

1.1 Massive Resource Integration, Laying the Knowledge Foundation

The ecological restoration industry is broad, encompassing ecological protection and restoration, watershed (river) management, coastal zone protection and restoration, ecological landscapes, and more. Each area involves vast amounts of technology, case studies, and policies/regulations. To build a comprehensive and rich knowledge base, we invested significant effort. Currently, the knowledge base includes 439 company-owned tendering/bidding entries, 519 planning & design entries, 161 engineering construction entries, 74 internal control management entries, and over 500 industry standards, totaling more than 10,000 organized project documents.

The knowledge base houses massive amounts of ZEHO ECO's proprietary project data.

1.2 Scientific Classification System, Precise RAG Strategy

First, based on business scenarios, we planned the top-level architecture of the knowledge base, defining key modules like tendering/bidding, planning & design, engineering construction, and internal control management, ensuring the knowledge system tightly aligns with actual business needs. Second, we hierarchically subdivided modules, constructing a tree-like classification structure for clear knowledge categorization. For example, within the engineering construction module, we segmented based on business scenarios: L1 main business (ecological environment), L2 sub-businesses (ecological engineering, ecological operation & maintenance), and L3 business products (river ecological restoration, urban brownfield restoration, nearshore marine ecosystem restoration, ecological seawall renovation, ecological forests, etc.), organized logically to lay the groundwork for later data vectorization. Finally, we organized actual knowledge resources (documents, data tables, R&D experimental data, etc.) according to this classification structure, creating corresponding sub-folders for orderly storage, enhancing knowledge reuse and management efficiency.

1.3 Advanced Technical Pathway, Building an Intelligent Closed-Loop Knowledge Governance System

We collected relevant data from various business domains, filtered, organized, and cleaned it to ensure quality and accuracy. High-dimensional vector technology was used to process the data, transforming unstructured knowledge into a vector format computers can understand and process. Concurrently, we built assistant workflows, conducted prompt engineering, and optimized inputs/outputs so models better understand user queries and needs. By fine-tuning model parameters like Temperature and top-p, we adapted them to different business scenarios and problem types, improving model generation capabilities and accuracy.

Knowledge Base Construction Technical Pathway

02 Scientific Evaluation System: 200-Point Precision Validation Across Three Dimensions, Rooted in Business, Applied to Scenarios

2.1 Rooted in Core Workflows and Pain Points of Four Business Segments

The technical team initially immersed itself in frontline operations. Through interviews, questionnaires, and analysis of actual work, they systematically identified the differentiated knowledge retrieval and application needs across segments. Planning & Design and Engineering Construction emphasize inspirational and innovative solutions; Tendering/Bidding and Internal Control Management demand precise, normative answers with extremely low error tolerance.

2.2 Scenario-Based Capability Validation Framework

To comprehensively measure whether the Z1+ Knowledge Base possesses the "scenario-specific intelligence" to solve real problems, the team built a "two-layer testing framework."

Layer One: Test Type Stratification (200 questions total) – Focus on Knowledge Application Forms

• Fact Query Type (80 items): Specifically tests the knowledge base's ability to locate and reproduce precise information (e.g., regulation text, dates, standards). Questions cover tender document details, key clauses of various contract templates, latest company internal control policies, ensuring "evidence-based" scenario needs are fully covered.

• Professional Query Type (120 items): Focuses on verifying the knowledge base's ability to understand, explain, synthesize, and apply professional knowledge (e.g., complex design specifications, construction process essentials, quality acceptance standards). Questions cover ecological design principles, special project acceptance procedures, compliance risk analysis, targeting professional scenarios requiring "deep understanding" and "strategic advice."

Layer Two: Quantitative Evaluation Metrics – Focus on Knowledge Quality

• Knowledge Coverage: Measures the ability to accurately associate and extract relevant information from the enterprise knowledge base composed of over 500 professional documents (Goal: No knowledge blind spots).

  
  

• Fact Consistency: Strictly compares errors between knowledge base outputs and original authoritative sources (document library), setting an enterprise-grade stringent standard: error rate must be <1.2% (Pursuing "zero errors").

  
  

• Logical Expressiveness: Evaluates if answers conform to professional vocabulary and logical expression in specific business scenarios, ensuring readability, professionalism, and logicality meet standards (96% matching rate).

  
  

• Establishing a "Control Group": To concretely demonstrate the significant enhancement value of the Z1+ Knowledge Base over the base large model (Z1) after "enterprise knowledge augmentation," a control group test was meticulously designed. All 200 test questions were answered simultaneously by the base Z1 model (not connected to the enterprise KB) and the Z1+ Knowledge Base. Horizontal comparison clearly revealed the knowledge base's "upgraded leap" in specific enterprise knowledge scenarios.

Tendering Test Set

Planning & Design Test Set

Engineering Construction Test Set

Internal Control Management Test Set

2.3 Business-Oriented Dynamic Intelligence Engine: Parameter Tuning On-Demand, Intelligent Adaptation

Quantifying Differential Needs & Strategy Formulation: Based on deep insights into segment needs, the technical team innovatively used the large model's "Temperature" parameter as the core adjustment lever. Through pre-test analysis and model behavior studies, differentiated parameter adjustment plans were precisely set:

• Planning & Design / Engineering Construction Segments: Temperature ↑30%. Repeated testing confirmed this increase effectively stimulates the model to generate diverse and forward-looking solution suggestions (e.g., technical routes, alternative construction methods, risk contingency plans) while maintaining baseline accuracy.

  
  

• Tendering/Bidding / Internal Control Management Segments: Temperature ↓40%. Extensive stability testing confirmed this reduction maximizes suppression of randomness, ensuring answers are highly focused on regulatory clause text and unique solutions, fulfilling the requirement of "following rules chapter and verse."

Actual Case Validation:

• Ecological Engineering Method Query (Planning & Design Scenario): The Z1+ Knowledge Base not only provided the standard solution but also extended recommendations for 2-3 new ecological engineering methods and their applicable conditions.

Planning & Design Scenario Q&A Test

Z1+ Knowledge Base Answer

Z1 Answer

• Internal Management Clause Query (Internal Control Scenario): Answers pinpointed the specific chapter (X), article (Y), and clause (Z) in the management system document, with a positioning error rate of 0.

Internal Control Scenario Q&A Test

Z1+ Knowledge Base Answer

Z1 Answer

2.4 Panoramic Acceptance: Deep Validation Across Three Dimensions

To ensure the professionalism and comprehensiveness of acceptance results, the company formed an acceptance group of 21 core members, including business leaders and IT experts. Using a rigorous 100-point scoring method combined with specific scenario requirements, they scored the Z1+ Knowledge Base and Z1 separately, ensuring objective, quantifiable results.

Acceptance Group Scores for Z1+ Knowledge Base and Z1

2.4.1 Knowledge Accuracy (Average Score: 91/100):

• Z1+ Knowledge Base Performance: Knowledge coverage - document positioning rate 99%, fact consistency error rate <1.2%, scenario matching rate 96%. Overall score: 91.

• Z1 Performance: Knowledge coverage error rate 28.5%, generalized expressions 74%. Overall score: 74.

2.4.2 Functional Completeness (Average Score: 92/100):

• Multi-turn Dialogue Retention: The group conducted high-intensity continuous questioning (e.g., 10 consecutive rounds), rigorously testing dialogue history memory, intent understanding, and contextual logical coherence, proving its practical multi-turn interaction capability.

2.4.3 Interaction Experience (Average Score: 85/100):

• Response Speed: Executing batch high-concurrency queries, test results showed 98% of queries returned results within 2 seconds, meeting efficient office demands.

• Dynamic Update Capability: To verify real-time updating and immediate effectiveness, the group imported the latest internal management document "Financial Invoice Management Manual" directly into the knowledge base system on-site. They then immediately retrieved and tested revised key clauses from the manual, achieving 100% retrieval accuracy.

Acceptance Scoring Details

03 Multi-party Evaluation: Knowledge Base Empowers Enterprise Development

Colleagues from the Planning & Design department stated the knowledge base provides rich design resources and inspiration. Designer Jin Guilong said: "Previously, projects required significant time to research materials and find case studies. Now, we can quickly find relevant design solutions and technical data. Human-machine interaction provides significant inspiration and innovation for projects, greatly shortening the design cycle. The knowledge base allows real-time access to the latest research and industry trends, continuously enhancing design foresight."

Tian Lei, Chief Engineer of the Project Management Center, praised the knowledge base: "During construction, the AI Large Model Assistant is like an experienced master, always providing technical support and solutions. When encountering construction challenges, searching relevant keywords in the knowledge base quickly yields similar cases and solutions, avoiding repeated trial-and-error, improving construction efficiency and quality. Simultaneously, the safety management knowledge ensures our construction is more standardized and safer."

Tan Yanyun, Head of the Tendering/Bidding Department, commented: "The tendering/bidding cases and strategies in the knowledge base provide crucial references for formulating bid proposals, helping us better understand tender document requirements and highlight the company's advantages and characteristics."

Tian Ying, Head of Internal Control Management, emphasized the knowledge base's vital role: "The knowledge base systematically organizes the company's systems and processes, facilitating employee queries and learning, improving policy execution efficiency, and providing strong support for compliant operations."

Building an enterprise knowledge base is an ongoing task requiring continuous investment and improvement. Practice has proven that the knowledge base not only enhances the enterprise's knowledge management and innovation capabilities but also provides powerful support across business segments, strengthening core competitiveness. In the future, we will continue optimizing the knowledge base's content and functionality to better serve enterprise development and contribute to the advancement of the ecological restoration industry.