Page 318 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Then, it initiates retrieval from the RAG (Retrieval-Augmented Generation) built on vertical
domain documents, and the retrieved content and extracted questions are fed into the
large language model, which generates precise and detailed recommendations according
to prompts, effectively improving the efficiency of marketing services and enhancing both
customer satisfaction and employees' labour productivity [1]. To further enhance the Retrieval-
Augmented capability, the RAG capability in this case adopts deep semantic encoding based
on the BGE (Embedding Model) to improve the generalization ability of knowledge retrieval.
It is pre-trained on a million-level exclusive dataset and dual-reinforced by supervised fine-
tuning and trainer feedback to train a specialized vector model for vertical domains. Through
vectorized retrieval, it achieves rapid data positioning and precise recall. It utilizes the LLM and
Prompt Engineering to optimize the generation strategy, improving the accuracy of intelligent
recommendations [3].

Currently, the RAG is built on a modular technology stack based on the LangChain framework.
Through the unified service interfaces of RESTful API and SDK, it has achieved standardized
implementation throughout the entire process, constructed a reusable enterprise-level solution,
and can be quickly replicated in other industries.

The self-reflection agent is applied in two primary scenarios, intelligent assistance and
intelligent quality inspection. In the context of intelligent assistance, it plays a critical role
in refining the output of a question summarization model. After analyzing the conversation
between a customer and a customer service representative, the model generates an initial
summary of the issue. This summary is then evaluated by the self-reflection agent, which
uses the original conversation as a benchmark. The agent assesses the summary based on
criteria such as accuracy, comprehensiveness, objectivity, and logical reasoning. If the agent
assigns a score below the predefined threshold, it identifies any deficiencies in the content and
generates targeted improvement suggestions. These suggestions, such as highlighting missing
information, are formulated into new prompts for the summarization model. The model then
regenerates the summary using this refined guidance. This evaluation and revision process is
repeated until the summary meets the quality threshold, ensuring that the final output is both
accurate and complete.

2�2 Benefits of the use case

In terms of “Decent Work”: The organization-wide adoption of AI not only helps service and
marketing personnel master new skills to adapt to the AI-infrastructure-driven employment
environment, reducing employment inequality among young people caused by the lack of new
technology training, but also creates emerging jobs such as AI data labeling. Meanwhile, agents
independently extract knowledge and solutions, complete quality inspection evaluations, and
update quality inspection report data, increasing the coverage rate of knowledge compilation
and quality inspection from 10% (manual) to 100%. Human resources from original positions
have transformed into AI annotators, assisting in providing feedback to the "intelligent
evolution engine" of meta-review agents, enabling continuous learning and self-evolution of
AI capabilities.

By integrating customer clustering analytics and historical behavioral data to build dynamic
customer profiles, our AI system combines semantic recognition and query reformulation
to accurately identify customer needs. Leveraging the Transformer architecture, it instantly
matches the most relevant knowledge documents or scenario-based solutions, delivering

282