Introduction
ImpReSS is a novel, LLM-powered recommender system designed specifically for customer support environments. Unlike traditional Conversational Recommender Systems (CRSs) that suggest items based on explicit user preferences (e.g., "Do you prefer horror or comedy movies?"), ImpReSS implicitly identifies what a user needs to solve a problem. In support contexts, users are focused on resolution rather than browsing, making traditional preference-gathering intrusive. ImpReSS bridges this gap by suggesting relevant products—such as a Wi-Fi extender for a user experiencing connectivity issues—without requiring explicit purchasing intent.
Methods: The ImpReSS Pipeline
The authors propose a three-step "add-on" pipeline for existing support chatbots:
- Query Generation: An LLM analyzes the conversation to produce a summary and diagnosis. It then generates preliminary Solution Product Categories (SPCs)—general product types that could address the root cause.
- Candidate Retrieval: The system queries specialized "Catalog Databases." These include indices populated via web searches for real-world use cases and LLM-generated feature descriptions. This multi-index approach ensures a diverse candidate set.
- Candidate Ranking (Bootstrap Approach): To mitigate "position bias" in LLMs, ImpReSS shuffles candidates and performs three ranking iterations, aggregating the results for a final prioritized list.
Evaluation and Datasets
ImpReSS was evaluated across three distinct domains:
- Cybersecurity Troubleshooting ($DS^{CT}$): 20 real-world technical support conversations.
- Information Security ($DS^{IS}$): 70 expert-level Q&A pairs from Stack Exchange.
- General Problem Solving ($DS^{GE}$): 224 synthetic conversations covering diverse topics like pet care and home maintenance.
Key Insights and Results
The system demonstrated high accuracy, achieving a Mean Reciprocal Rank (MRR@1) of 0.85 for $DS^{CT}$ and 0.82 for $DS^{IS}$. In the general dataset ($DS^{GE}$), it reached an MRR@1 of 0.72 and a Recall@3 of 0.89.
A critical finding was the necessity of the Bootstrap Ranking step. As shown in the performance metrics, the system's ability to correctly rank solutions improves significantly after the first iteration, overcoming the inherent biases of single-pass LLM ranking.
Furthermore, the study highlighted that retrieving based on "Use Cases" (web-derived data) was more effective than relying solely on product features. Performance also tended to improve as conversations progressed and more context became available for diagnosis.
Conclusion and Future Work
ImpReSS proves that effective recommendations can be made in support settings by focusing on problem diagnosis rather than user history. While promising, future work is needed to evaluate real-world conversion rates and optimize the timing of recommendations to ensure they remain helpful rather than intrusive. Additionally, exploring on-device deployment with smaller LLMs could reduce latency and operational costs.
Summary updated