The application practice of artificial intelligence based on the Nine-element vector database by the Information Center of a certain district in Shenzhen City

Case Summary

Relying on the Universal hyper-converged database product independently developed by Shenzhen Universal Database Co., LTD., the information center of a certain district in Shenzhen has built the country's first "AI+ Public Opinion Quick Service "full-stack domestic intelligent platform. By integrating innovative technologies such as spatio-temporal data modeling, multimodal semantic understanding, and group demand clustering, a closed-loop management of "second-level response, intelligent distribution, and virtual-real co-governance" for people's livelihood demands is achieved. The project innovatively adopts a hyper-converged architecture of vector data, spatio-temporal data, object data and relational data to develop nine core AI scenarios, reducing the average processing time of demands by 42.8% and lowering system resource consumption by 67%. Since its launch, the platform has handled a total of 368,000 demands, with a satisfaction rate of 98.6% among the public, forming a replicable model for digital reform in grassroots governance.

Implementation background

(1) The challenge of surging public demands: Before the renovation, the 12345 hotline in the entire district received over 300,000 cases. The traditional manual distribution mode led to 30% of the demands being processed overdue.

(2) System architecture bottleneck: The original MySQL database is unable to support an average of over 8,000 concurrent accesses per day, and the CPU usage rate exceeds 90% during peak periods.

(3) Data silo problem: 12 types of heterogeneous data such as video surveillance, GIS maps, and voice records cannot be effectively associated and analyzed.

(4) Localization requirements: It is necessary to achieve full-stack information technology innovation adaptation from chips, operating systems to databases.

Against this backdrop, the Information Center of this district, in collaboration with the Universal Database team, has created a new generation of AI-driven public welfare service platform based on the Universal hyper-converged Database.

Implementation goals

1.Build an intelligent hub: Establish a hyper-converged database that supports 1000-dimensional vector computing to achieve unified management of multimodal data.

2.Process reengineering: The entire process of handling demands has been compressed to within 3 working days, and the accuracy rate of AI distribution exceeds 95%.

3. Technological breakthroughs: Developed a semantic understanding model for demands (accuracy rate ≥92%) and a group event early warning system (early warning time ≤30 minutes).

4. Ecological Adaptation: Complete full-stack compatibility verification with domestic chips, operating systems, and databases.

Construction content

(I) Construction of hyper-converged database infrastructure

Hybrid data architecture

1.Deploy a cluster of nine hyper-converged databases, including the main database and two backup databases, to support high availability guarantee with RPO=0.

2. Innovatively adopt a four-dimensional data engine of "relational + vector + spatio-temporal + object" to achieve joint retrieval of text, voice and geographic coordinates.

Intelligent computing power scheduling

1.Build a dedicated computing node for AI inference, supporting direct connection of deepseek large models to the database;

2. Build a dynamic load balancing mechanism, with a request allocation delay of 50ms during peak periods;

(II) Development of core AI application scenarios

Intelligent input of demands to achieve digitalization of elements

1.Intelligent voice-to-text conversion enables citizens to complete the registration of their demands by simply saying one or a few sentences through the platform, and the system will automatically convert the voice into text.

2. Intelligently identify the elements of the appeal, recognizing the core elements such as the type of appeal, the address of the appeal, the person making the appeal, and the key words of the appeal contained in the appeal text;

3. Intelligently build an element database, automatically fill in the identified demand elements into the "digital form", and complete the digitalization of the demand elements.

Intelligent classification of demands to achieve data association

1.Identify sensitive requests by matching request elements against a sensitive word database, recognizing emergency, safety, and petition-related events within the request, with a focus on monitoring and pushing for them.

2.Identify group demands, automatically classify and identify group demands of "multiple people reflecting the same event", and handle them in a related manner.

3. Identify repeated demands. For repetitive demands that "have already been completed and are raised again", automatically associate them with historical work orders and prioritize their allocation to the original handlers for handling, ensuring the continuity and accuracy of the work.

Intelligent distribution of demands enables immediate response to complaints

Based on core data such as the list of powers and responsibilities, historical work orders, and "good or bad reviews", different weight indicators are set, and the top three disposal departments in terms of matching degree are recommended through a pre-trained classification model.

Intelligent analysis of demands, achieving a response to the people's calls

1.Intelligent follow-up visits are implemented to achieve "re-handling of negative reviews and satisfy the public". Obtain the evaluation results of citizens' satisfaction through intelligent outbound calls, and automatically return the unsatisfactory work orders for reprocessing.

2. Demand intelligent rendering to achieve "one-image presentation and interaction between virtual and real". Through the CIM spatio-temporal digital foundation, precise mapping and association of demand events are achieved. The AI early warning model establishes a ranking of hotspots in demand locations, forming a heat map of demand distribution to remind local streets and competent authorities to pay close attention. The intelligent video surveillance system realizes "virtual and real co-governance", promoting remote real-time handling of demands.

(III) Full-stack domestic adaptation

1.The performance optimization with Huawei Kunpeng 920 chip has been completed, and the tpmC index in the TPC-C test has been significantly improved.

2. Passed the compatibility certification of domestic operating systems such as UnionTech and Kylin;

3. Build national cryptographic algorithms, such as SM4 encrypted transmission channels, and the data storage encryption strength reaches AES-256;

Implementation effect

(I) Business process optimization

(II) Breakthroughs in technical performance

• The database query response time has been reduced from 1200ms to 280ms.

  
  

• The training efficiency of AI models has been increased by three times, and the retrieval of billions of vectors takes only one second.

  
  

• Hardware resource utilization optimization: CPU peak load dropped from 92% to 28%, and memory usage decreased by 64%.

(III) Remarkable social benefits

1.Centering on four major links - intelligent input, intelligent classification, intelligent distribution, and intelligent analysis - nine application scenarios will be created to complete the construction of a smart platform that is user-friendly for the public, willing for the government, and accessible to society, and form a "four-step closed loop, full-process intelligent" demand handling mechanism.

2. The total number of work orders, the on-time completion rate, the average processing time, and the overall satisfaction rate all rank among the top in all districts of the city.

Innovation highlights

Hyper-converged database architecture

The world's first "relational + vector + spatio-temporal + object" four-engine integrated design, with a single database supporting compound operations such as SQL query, vector similarity calculation, and geofencing detection;

Dynamic weight distribution model

Innovatively proposed the "Departmental Capability Value" algorithm, which intelligently recommends disposal units by comprehensively considering the historical completion rate (with a weight of 40%), the public approval rate (35%), and the professional matching degree (25%).

The governance model of virtual and real interaction

The real-time mapping between the digital twin city and the physical world is achieved through the spatio-temporal data processing capabilities of the Universal hyper-converged database.

Experience summary

The successful implementation of this project proves:

1.Vector databases have demonstrated significant advantages in the processing of unstructured data, improving the accuracy of semantic understanding by 23% compared to traditional solutions.

2. Through the deep coupling of "AI+ database ", the handling of people's livelihood demands can shift from experience-driven to data-driven.

3. The full-stack domestic solution fully meets the performance requirements of the government affairs system, and the construction and operation and maintenance costs are reduced by 40% compared with expectations.

This model has been expanded to 15 fields including emergency management and market supervision, and it is expected to form a digital ecosystem for urban governance worth over 10 billion yuan within three years.

- Reprinted from 'Informationization Observation Network