From Third‑Party to Self‑Developed: Zhaozhuan’s AI Chatbot Evolution

1 Introduction

After launching in 2015, Zhaozhuan’s robot customer‑service system went through three major version upgrades. The core NLP capability relied on external vendors, limiting breakthroughs.

In 2024, with mature AI technology and resources, the company intensified AI investment, making the customer‑service robot a key focus. Leveraging accumulated business and system data, the MOSS system achieved fully self‑developed core NLP, eliminating dependence on external vendors and dramatically improving intelligent service.

2 Evolution: From Concept to Implementation

2.1 Choosing the Technical Path – NLP or AIGC?

Most enterprise chatbots still use mature, engineering‑ready NLP as the foundation. Our decision was based on three key considerations:

Strong controllability: Customer service requires precise, explainable responses; large language models can produce hallucinations and uncontrolled outputs.

Low cost and efficiency: Training an NLP module typically needs only a few thousand labeled examples, far cheaper than fine‑tuning large language models.

Immediate maintainability: Knowledge‑base updates can be made quickly via a configuration platform, ensuring agile service.

Although AIGC shows strong capabilities in open‑ended dialogue, its issues with hallucination, high resource consumption, and delayed knowledge updates made us stick with mature NLP to meet the strict rules and high accuracy required by Zhaozhuan’s service.

For ambiguous user inputs (e.g., “okay”, “repeat”, “what next”) and out‑of‑scope questions, we introduced a Retrieval‑Augmented Generation (RAG) model. RAG rewrites the query using conversation context, retrieves relevant internal knowledge, and generates intent‑aligned answers, greatly enhancing contextual understanding.

2.2 Model Construction and Training – Building on a Solid Base

The customer‑service domain naturally accumulates massive structured knowledge and rich interaction logs, providing a solid foundation for a self‑developed model.

We have amassed a large number of standard Q&A entries and similar‑question data. After extensive verification, we selected the industry‑proven pre‑training model BERT to build the core knowledge vector library.

During training, we leveraged millions of high‑quality semantic recognition data collected from years of third‑party service usage. These data served as valuable “nutrition” for our model, enabling rapid training and parameter tuning. Combined with precise annotator corrections and targeted optimization of online bad cases, our self‑developed NLP model quickly surpassed the performance of the previous vendor model.

To continuously improve service quality and expand knowledge coverage, we designed a dynamic knowledge‑extraction workflow. Large models analyze robot‑human conversation logs, extract potential new knowledge suggestions, and are then reviewed and refined by annotators, ensuring the knowledge base remains up‑to‑date and comprehensive.

2.3 Steady Deployment – Online Gray‑Scale Testing and Data Validation

Before full launch, we performed rigorous validation. We first extracted a batch of newly generated real user queries and conducted extensive offline “mock exams” on the self‑developed model. Only after meeting performance thresholds did we deploy it to production.

We also upgraded the system architecture to support flexible NLP model configuration and traffic splitting, enabling A/B testing of different models and rapid rollback if issues arise, ensuring zero‑downtime risk control.

Accurate identification: Upon a user’s inquiry, the robot quickly matches the query using the self‑developed NLP model.

Data fusion: Based on the identification result, the robot dynamically calls various business‑system APIs to retrieve precise order, service status, and other data.

Intelligent generation and adaptation: Combining business data and scenario, the system generates the best answer template, fills placeholders, configures self‑service buttons, and applies hand‑off rules for personalized responses.

Handling vague or unknown inputs: For ambiguous replies or unclear semantics, the system invokes the RAG engine to supplement intent and re‑match; for questions beyond NLP coverage, RAG integrates knowledge‑base information to generate replies.

Knowledge‑base extraction and training: Continuous analysis of historical conversations by large models extracts new knowledge, constantly enriching the robot’s knowledge base.

3 Conclusion

Transitioning from third‑party reliance to a fully self‑developed solution marks a major technological leap and a strategic milestone for autonomous control. By carefully selecting NLP as the core and augmenting it with RAG, leveraging deep customer‑service data, efficient training, and a dynamic knowledge‑base system, we built a high‑precision, controllable, and easily maintainable intelligent service foundation.

Looking ahead, continued investment in self‑research capabilities and integration of cutting‑edge AI will further evolve Zhaozhuan’s chatbot, enhancing user experience and service efficiency, and providing robust technical support for the company’s high‑quality growth.