The previous article introduced the three-dimensional methodology of intelligent customer service robot demand research, and this article talks about the process of intelligent customer service evolving from relying on traditional knowledge bases to systems with intelligent “brains”. This shift reflects the advancement of technology and the pursuit of superior performance and user experience in the field of intelligent customer service. This article will discuss how to create an efficient and intelligent customer service robot system from the aspects of knowledge base construction and robot operation optimization, provide practical solutions for enterprises, and promote the development of intelligent customer service robots in a more intelligent and personalized direction.
1. Knowledge base construction: create the intelligent brain of intelligent customer service robots
1.1 Determine the scope of the knowledge base and organize materials
The picture comes from the Internet
The first step in building a knowledge base is to clarify its scope, which includes organizing materials, refining questions, judging knowledge forms, and writing answers.
Take e-commerce platforms as an example:
- Material collation can include user consultation records, product information, promotional activities, etc.;
- problem refinement can focus on high-frequency questions consulted by users;
- Knowledge form judgment can determine the answer form of the question (such as text, picture, video, etc.).
1.2 Improve semantic parsing capabilities
Semantic parsing capabilities are one of the core competitiveness of intelligent customer service robots.
To improve this capability, we need to optimize the matching logic of the knowledge base and the writing specifications of similar questions.
Taking bank customer service as an example, the matching rate of the robot can be improved by increasing the number of similar questions and optimizing the matching algorithm. By clarifying the writing specifications for similar questions, it can ensure that the bot can accurately understand user intent and provide effective answers.
1.3 Knowledge base construction and test tuning
The picture comes from the Internet
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
After the knowledge base scope is determined and the semantic parsing ability is improved, we can enter the knowledge base construction and test tuning stage. This stage includes FAQ category construction and problem import, basic function configuration, thesaurus configuration, and knowledge base testing and tuning.
Take e-commerce platforms as an example:
- FAQ category construction can be carried out around themes such as commodity consultation, order inquiry, and after-sales service.
- The basic function configuration can include welcome messages, navigation menus, manual buttons, etc.;
- Thesaurus configuration can include keywords such as product name and brand name.
2. Robot operation optimization strategy: from launch to continuous improvement
2.1 Bot launch strategy and risk control
The launch of the robot is an important milestone in the intelligent customer service robot service.
Before going live, we need to develop a detailed launch strategy and control risks.
Grayscale publishing is an effective launch strategy that allows us to test on a small number of users or channels, observe the online performance of the bot, and collect user feedback.
Taking e-commerce platforms as an example, Grayscale Publishing can first launch intelligent customer service robots on some product pages or user groups, and decide whether to launch it in full based on the test results.
2.2 Robot operation indicators and element management
Robot operations need to pay attention to the two core indicators of matching rate and resolution rate, as well as three elements: standard questions, similar questions and answers.
Take bank customer service as an example:
- Improve the matching rate by optimizing the matching logic of the knowledge base and writing specifications for similar problems;
- Improve the resolution rate by optimizing the content of answers and improving the effectiveness of answer display.
- At the same time, refined management of standard questions, similar questions and answers ensures the accuracy and effectiveness of the knowledge base.
2.3 Resolution rate improvement strategy and continuous optimization
The picture comes from the Internet
For the problem of low resolution rate, we can adopt a variety of strategies to improve.
For example, timely supplement knowledge points for the problem of lack of knowledge;
- Optimize matching logic or write specifications for similar problems for matching error problems;
- set unreasonable questions for answers, re-examine the content of answers, and optimize the display effect;
- Set up unreasonable human entry policies and adjust policies to ensure that users can get timely human help when needed.
Taking e-commerce platforms as an example, by continuously optimizing these strategies, the robot’s resolution rate and user experience can be continuously improved.
3. Decision-making layer technical solutions for intelligent customer service robots
The technical solutions of intelligent customer service robots have undergone an evolution process from tradition to modernity, which not only reflects the progress of natural language processing technology, but also reflects the unremitting pursuit of higher performance and better user experience.
In traditional scenarios, the technical implementation of intelligent agent robots mainly relies on a well-built knowledge base and matching methods based on rules or templates. These solutions distilled high-frequency questions and their answers by sorting out customer service experience, historical conversations, and industry knowledge, forming a question-and-answer (FAQ) format. However, this solution has obvious limitations, such as high maintenance costs of knowledge bases, weak generalization capabilities, and difficulty in coping with complex and changeable user needs.
With the rapid development of natural language processing technology, the technical solutions of intelligent customer service robots have ushered in major changes. The emergence of pre-trained models, such as BERT and GPT, provides intelligent customer service robots with powerful language understanding and generation capabilities. These models have mastered rich linguistic knowledge and patterns through unsupervised learning on large-scale corpora, which can better understand the semantics of user problems, significantly improving the accuracy and generalization ability of semantic analysis.
However, pre-trained models, while powerful, may still face challenges when applied directly to specific domains. Therefore, the industry has further explored the technical solution of pre-trained model combined with SFT (Supervised Fine-Tuning) in the special field. SFT in the special field introduces domain-specific corpus and annotation data to supervise fine-tuning the pre-trained model, so that the model can better adapt to the language habits and business rules in the field. This solution not only retains the generalization ability of the pre-trained model, but also significantly improves the professionalism and accuracy of the model in specific fields.
At present, the mainstream technical solution of intelligent customer service robots is based on pre-trained models + SFT in the special field, and further introduces RAG (Retrieval-Augmented Generation) technology. RAG technology combines the advantages of information retrieval and generative models, enabling dynamic retrieval of relevant information from external knowledge sources as responses are generated. This allows the bot to handle more complex and open-ended questions while maintaining accuracy and relevance in its answers.
Through RAG technology, intelligent customer service robots can not only leverage the powerful capabilities provided by pre-trained models and specialized field SFTs, but also connect to external knowledge bases in real time to ensure the timeliness and accuracy of answers. The technical solution of intelligent customer service robots has undergone an evolution process from traditional knowledge base matching to pre-trained model + SFT in special fields, and then to the current mainstream solution (pre-trained model + SFT + RAG in special fields).
This process not only reflects technological advancements but also reflects the industry’s relentless pursuit of higher performance and better user experience.
In the decision-making layer technical solution of intelligent customer service robots, the combination of pretrained model (Pretrained Model) + special field SFT (Supervised Fine-Tuning) + RAG (Retrieval-Augmented Generation) is the most mainstream and efficient architectural design at present. This combination realizes the deep adaptation of model capabilities and business requirements through hierarchical collaboration, which can be broken down as follows:
3.1 Analysis of the three-layer architecture of technical solutions
Pretrained Model
Role: Provide a base for general language understanding and generation capabilities (such as GPT, LLaMA, ChatGLM, etc.).
Advantages: Learn general grammar, semantics and knowledge expression through massive unannotated data, and have strong generalization ability.
Positioning: The “general brain” of the decision-making layer, handling basic semantic analysis and contextual modeling.
Special Areas SFT (Supervised Fine-Tuning)
Function: On the basis of the pre-trained model, use domain annotation data to fine-tune model parameters and inject vertical domain knowledge.
Mission Critical: The church model follows industry terminology, business norms (e.g., medical terminology, legal provisions); Optimize the output style (such as politeness and conciseness of customer service speech); Align user intent (e.g., returns in e-commerce scenarios, multi-intent identification of complaints).
Data requirements: High-quality field instructions-answer pairs are required (e.g., {“instruction”:”Return process”, “output”:”Log in to your account→ submit an application→ mail the item”}).
RAG（Retrieval-Augmented Generation）
Function: Dynamically access to the external knowledge base to solve the problem of “knowledge solidification” of the model.
Operation process: Retrieval: retrieve relevant document fragments from the vector database according to user questions; Enhancement: Stitch search results into input prompts; Generation: The model generates a final response based on the search content.
Core values: support real-time knowledge updates (such as policy changes, product parameters); Reduce hallucinations (answers from authoritative knowledge bases); Highly interpretable (user traceable source of answers).
3.2 Synergy logic of three-layer technology
Examples of collaborations:
User asked, “What is the new policy for medical insurance reimbursement in 2025?”
RAG retrieves the latest policy documents → Pre-trained/SFT model Understands questions and generates answers that match medical discourse → Outputs “According to the new regulations in 2025, the outpatient reimbursement ratio will be increased to 70% (policy source: Health Insurance Bureau Document 2025-001)”.
3.3 Technical selection suggestions
Pre-trained model selection
- General scenarios: Prioritize models with strong multilingual support (such as GPT-4, DeepSeek);
- Chinese scenarios: Domestic models such as ChatGLM3, Qwen, and Baichuan are more suitable for localized expressions.
SFT optimization focus
- Data quality: avoid low-quality labeling, and it is recommended to use domain experts to review data;
- Robustness training: Add typos and mixed Chinese and English prompts to improve anti-interference ability;
- Task tag management: Classify data by scenario (such as Return Process and Troubleshooting).
RAG Engineering Practice
- Knowledge base construction: PDF/HTML parsing→ text chunking → vectorization (embedding model selection such as BGE and text2vec);
- Search strategy: hybrid BM25 (keyword) + semantic retrieval, balancing accuracy and recall;
- Exception handling: Design reply templates for empty and contradictory searches (e.g., “No relevant information found, please contact manual customer service”).
3.4 Typical application scenarios
summary
Pre-trained model + domain SFT + RAG It is the golden technology combination of the decision-making layer:
- The pre-trained model provides a “general intelligent base”;
- SFT realizes “field specialization transformation”;
- RAG gives “dynamic knowledge expansion capabilities”.
The three synergize to form a closed loop, which not only ensures the controllability of core capabilities, but also solves the timeliness of knowledge and personalized needs, and has become the standard architecture of enterprise-level intelligent customer service.
4. Requirements development and management process
4.1 Requirements development
Requirements development includes two stages: user requirements research and product requirements definition.
- In the user demand research stage, we dig deeper into user needs through 5W2H models and other methods.
- In the product requirements definition stage, we transform user requirements into specific product requirements to provide a basis for subsequent product design and development.
Taking intelligent customer service robots as an example, product requirements can include functional requirements, performance requirements, security requirements, etc.
4.2 Requirements management
Requirements management includes processes such as requirements confirmation, review, tracking, and change control.
- Through requirements confirmation, we ensure that the development team has a common understanding of the requirements;
- Through requirements review, we identify and solve potential problems in requirements;
- Through requirements tracking, we ensure that development results are aligned with design goals;
- Through requirements change control, we manage the impact of requirements changes and ensure the smooth progress of the project.
Taking e-commerce platforms as an example, demand management can focus on the priority and progress of user inquiries.
5. The importance of critical thinking ability and field research
5.1 Strengthen critical thinking
In the development process of intelligent service products, we need to strengthen our critical thinking ability to avoid falling into thinking traps. For example, we can’t simply think that AI will completely disrupt human services, but should see human-machine coupling as the best solution.
AI has found a new balance between cost and service experience for businesses, but human services still have irreplaceable advantages in some aspects.
5.2 Field research
Field research is an important means to understand the real needs of users.
Through field research, we can find the real cause of the problem, not just imagination or superficial user feedback.
Field research has a significant impact on robot training after launch, helping to solve the problem of customers being reluctant to use online robots.
Taking bank customer service as an example, through field research, we can have a deeper understanding of users’ usage habits and demand pain points, and provide strong support for the continuous optimization of products.
6. Conclusion
This time, through the first and second articles, we briefly explored the intelligent customer service robot from“Demand Research” To “knowledge base driven” and then to “smart brain” The process:
- Knowledge engineering: Build an accurate knowledge base through user question classification and material cleaning Semantic analysis ability leaps；
- Operate dual engines: Built with matching rate and resolution rate as the core indicators, combined with grayscale release and element management Continue to optimize the closed loop；
- The golden triangle of technology: Pre-trained model + field SFT+RAG architecture becomes Industry standard configuration scheme(e.g., medical/financial scenario verification);
- Demand-driven: Control the whole process from development to management to ensure that the robot responds to fit the business evolution.
The next stage of breakthroughs in intelligent customer service will focus:
- Dynamic knowledge flowDeep coupling of RAG with real-time business systems (e.g., direct connection of order/inventory data);
- Emotional interaction: Adaptive speech generation based on emotion recognition;
- A new paradigm of human-machine collaboration: The intention relay mechanism between human agents and AI.
The great changes brought about by the development of science and technology are inevitable, and how to use new technologies to connect to current work and applications requires us to continue to study. I look forward to discussing intelligent customer service robots with you~
