Changan Government AI Assistant: Implementation and Evolution of a Large Language Model-Based Intelligent Service System

Technical Background and Objectives

The digital transformation of government services has made intelligent service delivery a critical priority. Traditional customer service systems face three major challenges: outdated knowledge bases with accuracy rates below 60%, inability to handle complex business scenarios through standardized Q&A, and limited service hours affecting citizen experience. Changan District Government Service Center introduced an “AI+Government” innovation strategy by building a large language model-powered intelligent service system with three core objectives:

1. 24/7 Service Coverage: Breaking traditional working hour limitations
2. End-to-End Process Integration: Extending from consultation to material pre-review and form assistance
3. Multi-Scenario Adaptation: Covering 339 high-frequency services with multi-turn conversation capabilities

The system adopted a two-phase evolution strategy: launching a basic intelligent customer service platform initially, then upgrading to a digital avatar-enabled government assistant through technology integration, forming a three-tier capability system from basic Q&A to intelligent guidance and full-process assistance.

System Architecture and Technical Implementation

Core Architecture Design

The microservices-based system comprises four key modules:

• Large Model Service Layer: Deployed a 100B-parameter model specialized for government services, supporting intent recognition, entity extraction, and multi-turn conversation management
• Knowledge Middle Platform: Structured knowledge graph containing 2,800+ Q&A pairs and 1,500+ material templates across 339 service items
• Business Adaptation Layer: Developed specialized plugins for material format validation and certificate verification
• Multi-Channel Interaction Layer: Integrated web, mobile, and self-service terminals supporting voice, text, and mixed-media interactions

Key Technical Innovations

Multi-Stage Training Optimization:

• Base training: Pre-trained on general corpora for foundational language understanding
• Domain adaptation: Enhanced with 200,000 government dialogue samples through continued training
• Reinforcement learning: Implemented PPO algorithm for response strategy optimization using user satisfaction feedback
• Continuous learning: Established incremental learning pipeline updating 300+ new knowledge items weekly

Digital Avatar Integration:
The second-phase upgrade introduced “Chang Xiao’an Digital Avatar” through 3D modeling, speech synthesis, and motion control technologies. Implementation example:

# Digital Avatar Engine Framework
class DigitalAvatarEngine:
    def __init__(self):
        self.nlp_engine = LLMProcessor()  # Core LLM processor
        self.tts_module = VoiceSynthesizer()  # Text-to-speech
        self.animation_ctrl = MotionController()  # Motion generation
    def process_request(self, user_input):
        # Multimodal understanding
        intent, entities = self.nlp_engine.analyze(user_input)
        # Business processing
        response_data = self.business_handler(intent, entities)
        # Multimodal generation
        audio = self.tts_module.generate(response_data['text'])
        animation = self.animation_ctrl.generate(response_data['emotion'])
        return {
            'audio_stream': audio,
            'animation_params': animation
        }

Functional Evolution and Capability Upgrades

The system developed through two key phases:

Phase 1: Basic Intelligent Customer Service

Launched in February 2025 with three core functions:

1. Intelligent Q&A: Covering 85% of common scenarios with 92% accuracy
2. Material Guidance: Automatic generation of material checklists with PDF/Word download support
3. Process Navigation: Visual flowchart display of service procedures

Sample interaction:

User: What materials are needed for a food business license?
System: Based on your business type, please prepare:
1. Business license copy
2. Premises proof (property deed or lease contract)
3. Food safety management document
...(8 items listed with template download links)

Phase 2: Full-Process Assistance

The July 2026 digital avatar upgrade introduced four enhanced capabilities:

1. Material Pre-review: Automatic validation using OCR+NLP technologies
2. Form Auto-fill: Voice-controlled application form completion
3. Progress Tracking: Real-time status updates through government middleware integration
4. Exception Handling: Rejection reason identification and correction suggestions

Post-upgrade performance in road transport business license processing:

• Average handling time reduced from 45 to 12 minutes
• First-pass success rate improved to 89%

Operational Framework and Continuous Optimization

A three-dimensional operational mechanism ensures system effectiveness:

Knowledge Base Management

• Collection: User logs, agent feedback, and policy document analysis
• Processing: “Manual review + automatic annotation” quality control
• Application: A/B testing for response strategy optimization

Performance Monitoring

12-metric evaluation model:
| Category | Metrics | Target |
|————————|—————————————-|————-|
| Response Quality | Accuracy Rate | ≥95% |
| Service Efficiency | Average Response Time | ≤1.5s |
| User Satisfaction | NPS Score | ≥80 |
| Business Coverage | Service Item Coverage | 100% |

Iterative Optimization

Closed-loop process:

1. High-frequency issue identification through log analysis
2. Knowledge gap localization via clustering analysis
3. Model enhancement through specialized training
4. Optimization validation via grayscale release

Implementation Results and Industry Value

The system achieved significant improvements:

• Service Efficiency: Daily consultations increased from 300 to 1,200 with 65% reduction in agent workload
• User Experience: Satisfaction score improved from 82 to 94 (out of 100)
• Management Upgrade: Established replicable government knowledge management methodology with knowledge update cycle shortened from monthly to weekly

The practice provides three paradigm values for government digitalization:

1. Technology Integration: Demonstrated协同 application path of large models and digital avatars
2. Service Innovation: Constructed end-to-end service chain from consultation to assistance
3. Operational Management: Established knowledge-driven continuous optimization mechanism

The system is now expanding to street-level communities, targeting full coverage of 16 sub-districts and 232 communities by 2027 while exploring deeper integration with urban management platforms and IoT systems to continuously expand the boundaries of government intelligence applications.