Cleo: A Transparent and Controllable Chatbot for Conversational Commerce

Overview

Cleo is a conversational product advisory system designed to address opacity and unpredictability in large language model-based commerce applications. The system combines explicit ranking mechanisms with constrained language generation to enable transparent and controllable product recommendations. The architecture separates deterministic product ranking from natural language generation, applying categorical filters and numeric loss functions across 3,638 product specifications while constraining the LLM output to grounded catalog evidence.

Methods and approach

The system employs a hybrid architecture with two primary components. The ranker applies deterministic categorical filtering and numeric loss functions to product specifications, generating auditable per-attribute loss values that explain ranking decisions. The language generation component uses a constrained LLM that operates within boundaries defined by catalog data, reducing hallucination risk and ensuring descriptions remain grounded in available product information. The system implements natural language comparison generation and highlights functionality to support user decision-making. Information needs elicitation occurs through conversational refinement with real-time re-ranking capabilities, allowing users to inspect loss explanations and receive multi-item comparisons.

Key Findings

The demonstration system successfully integrates conversational interaction with algorithmic transparency and controllability. The architecture enables fluid conversation while maintaining visibility into ranking decisions through per-attribute loss explanations. The constrained LLM generation approach prevents hallucinated or persuasive content generation while maintaining naturalness in product descriptions. The system supports iterative refinement through real-time re-ranking and provides decision support through comparative features.

Implications

The approach advances the design of transparent and controllable conversational systems for product search and recommendation, addressing limitations of both traditional faceted search interfaces and opaque LLM-only recommenders. By separating deterministic ranking from language generation, the system maintains interpretability while enabling natural conversational interaction. The work provides a methodological framework applicable to conversational search and recommendation domains where transparency and user control are essential design requirements.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.