2026-07-28T23:36:27Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:26713 2026-07-28T09:32:41Z ddc:004 open_access

A Silent Multi-Agent Recommendation Engine with Graph-Augmented Memory for Immersive B2B E-Commerce Eleuterio, Davi Silva Oliveira, Pedro Filipe Matos, Paulo multi-agent systems agentic RAG product recommendation LangGraph Currently, e-commerce platforms are transitioning from static 2D catalogues to immersive environments, requiring recommendation systems to process complex behavioural signals. Traditional collaborative filtering struggles with cold-start scenarios and lacks semantic understanding of product catalogues. This paper presents a silent multi-agent recommendation engine developed for the VIMOS (Virtual Integrated Market Online Shopping) project, designed to operate without explicit textual input. Orchestrated via a LangGraph Directed Acyclic Graph, the system utilises Agentic Retrieval-Augmented Generation (RAG) to dynamically synthesise behavioural signals, dense vector retrieval (Qdrant), and knowledge graph traversal (Apache AGE). A locally hosted Llama 3.1 8B model autonomously extracts implicit user preferences to update a persistent graph database and generates deterministic, human-readable explanations for each recommendation. A two-shot verifier engineering pattern is introduced to resolve the structural incompatibility between dynamic candidate sets and static graph execution. The architecture is validated using a controlled synthetic dataset of B2B industrial products - sufficient for functional validation of all agent routing paths - demonstrating high retrieval accuracy, strict constraint adherence (stock and availability verification), and sub-4-second end-to-end latency (when the generative semantic justification step is bypassed) with no external API dependencies. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Davi Silva Eleuterio and Pedro Filipe Oliveira and Paulo Matos 2026 Is Part Of OASIcs, Volume 144, 15th Symposium on Languages, Applications and Technologies (SLATE 2026) InProceedings Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/OASIcs.SLATE.2026.15 urn:nbn:de:0030-drops-267131 https://drops.dagstuhl.de/entities/document/10.4230/OASIcs.SLATE.2026.15 eng https://creativecommons.org/licenses/by/4.0/legalcode