OlaMind: Towards Human-Like and Hallucination-Safe Customer Service for Retrieval-Augmented Dialogue

TL;DR

OlaMind tackles hallucination and lack of human-likeness in retrieval-augmented customer service by introducing a two-stage framework: Learn-to-Think, which distills structured CSR reasoning and response strategies, and Learn-to-Respond, which iteratively refines cold-start SFT and RL in a basic-to-hard progression. A hallucination-judge system combines human and LLM feedback to detect and mitigate non-factual content, guiding reward design and data refinement. Offline results show OlaMind-Hard-RL achieving higher human-likeness, better dialogue quality, and stronger safety, while online A/B testing demonstrates substantial improvements in intelligent resolution rate and reductions in human takeover rate across community-support and livestream scenarios. The approach offers a scalable, generalizable path for deploying hallucination-safe, human-like RAG-based dialogue in real-world customer-service settings, with publicly available code and data planned. Future work will extend instruction following, complex reasoning, and multi-agent system applications.

Abstract

Intelligent customer service (ICS) systems via retrieval-augmented generation (RAG) have been widely adopted in Web-based domains such as social platforms and e-commerce, achieving remarkable improvements in automation and efficiency. However, notable limitations still remain: these systems are prone to hallucinations and often generate rigid, mechanical responses, which can introduce business risks and undermine user experience, especially in Web-based customer service interactions under the RAG scenarios. In this paper, we introduce OlaMind, a human-like and hallucination-safe customer service framework for retrieval-augmented dialogue. Specifically, it first leverages a Learn-to-Think stage to learn the reasoning processes and response strategies from human experts, and then employs a Learn-to-Respond stage to perform cold-start supervised fine-tuning (SFT) combined with reinforcement learning (RL) for basic-to-hard self-refinement. Our method significantly enhances human-likeness and naturalness while effectively mitigating hallucinations and critical business risks. We have conducted large-scale online A/B experiments in an industry-level social customer service setting, and extensive experimental results show that OlaMind achieves significant cumulative relative improvements with intelligent resolution rates +28.92%/+18.42% and human takeover rate -6.08%/-7.12% in community-support/livestream-interaction scenarios, respectively, which highlights its consistent effectiveness across diverse real-world applications. The code and data will be publicly available.

Figures (7)

• Figure 1: Comparisons between human customer service, intelligent customer service with general LLMs, and our proposed OlaMind. Notably, OlaMind can think and respond with human-likeness and hallucination-safety, thus significantly bridging the gap.
• Figure 2: The overall framework of OlaMind. It introduces Learn-to-Think to focus on learning reasoning processes and response strategies from humans, and implements Learn-to-Respond with iterative refinement of cold-start SFT and RL in a basic-to-hard order, achieving human-like and hallucination-safe intelligent customer service.
• Figure 3: Comparison of the reasoning process and response strategy between OlaMind-Think and general LLMs.
• Figure 5: Performances across different stages of OlaMind.
• Figure 6: Comparisons of human-likeness score distribution.