Infobip Launches Agentic CRS: Conversational Recommender System for Live E-commerce Catalogues

Summary:Infobip has launched Agentic CRS, a conversational recommender system designed to handle live updates in e-commerce catalogues. The system employs a self-refreshing retriever to process changes in product feeds, synchronizing only the delta into a vector index to avoid full catalogue rebuilds. The core of the system is a controller-based dialogue layer that leverages an LLM solely for intent classification and preference elicitation, while retrieval, reranking, and diversity selection are handle

---

Core Breakthrough

Infobip's Agentic CRS is a conversational recommender system tailored for live e-commerce catalogues, featuring the following key technical highlights:

• Self-Refreshing Retriever: The system identifies changes in product feeds (additions, modifications, deletions, or unchanged items) and synchronizes only the delta into the vector index, avoiding the overhead of full catalogue rebuilds.
• Controller-Based Dialogue Layer: The dialogue layer is driven by a controller architecture that invokes the Large Language Model (LLM) solely for intent classification and preference elicitation, while retrieval, reranking, and diversity selection are handled by dedicated functions, enhancing system efficiency.
• Multi-Turn Dialogue Support: The system supports multi-turn interactions, providing accurate product recommendations in dynamic e-commerce environments.

Technical Analysis

The core component of Agentic CRS is a self-refreshing retriever that operates through the following steps:

1. Data Ingestion and Record Enrichment: It ingests data from merchant product feeds and enriches the product records.
2. Incremental Synchronization: By using per-item hashes, it identifies changes and synchronizes only the delta into the vector index.
3. Dialogue Layer Processing: The controller-based dialogue layer uses the LLM for intent classification and preference elicitation, while retrieval, reranking, and diversity selection are handled by dedicated functions for improved efficiency.

Industry Impact

The release of Agentic CRS provides e-commerce platforms with an efficient solution for handling live catalogue updates, which can enhance user experience and conversion rates. Its design principles also offer insights for other AI systems that need to manage dynamic data. Additionally, the system's deployment as a WhatsApp shopping assistant demonstrates its potential in instant messaging platforms.

Developer Recommendations

• Focus on Real-Time Data Processing: Developers can learn from Agentic CRS's self-refreshing retriever to optimize their systems' ability to handle real-time data.
• Optimize Dialogue Layer Design: By separating LLM invocation and dedicated function processing, developers can improve the efficiency and response speed of their dialogue systems.
• Explore Multi-Platform Applications: Consider applying similar technologies to other instant messaging platforms or intelligent assistants to expand application scenarios.

---

Source: ArXiv cs.IR (2026-08-27)

— END —

Tags: #Infobip #Conversational Recommender System #E-commerce AI #LLM Application #Real-Time Data Processing