Percorrer por autor "Bispo, Ivo Afonso Freire"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Retrieval-Augmented Generation for B2B Spare Parts Search Design and Evaluation of a Hybrid Retrieval and LLM-Ranking Recommendation APIPublication . Bispo, Ivo Afonso Freire; Craveiro, Olga Marina FreitasIndustrial procurement increasingly relies on digital product catalogs, yet non-specialist personnel often struggle to locate spare parts when they lack precise technical terminology or catalog identifiers. This vocabulary mismatch between natural-language descriptions and structured product specifications limits the effectiveness of traditional keyword-based search. This internship report was developed in the scope of an internship at Valtech, a consultancy company specialising in B2B software products. The work addresses a practical challenge faced by manufacturers: enabling non-specialist personnel to search for spare parts for industrial machinery using natural language, without requiring knowledge of precise catalog identifiers or technical terminology. This internship report presents the design, implementation, and evaluation of an AI-driven product search and recommendation API for B2B spare parts. The system applies a Retrieval-Augmented Generation (RAG) pipeline in which a hybrid retriever (dense semantic similarity + BM25 lexical retrieval) retrieves candidate products from a structured catalog, and a large language model reranks candidates and generates structured recommendations. Using a proof-of-concept catalog of 50 products and a 50-query test suite, systematic evaluation identified a semantic-dominant hybrid configuration (70:30 semanticto- keyword) that improved retrieval performance by 15 percentage points over pure semantic retrieval. The end-to-end pipeline achieved an 88% pass rate, exceeding the predefined 85% quality threshold, where a test is considered a pass when an expected SKU appears in the recommendation list. To address latency dominated by model inference, a query-level response cache reduced response time for repeated queries by approximately 97%. The results provide evidence that RAG-style architectures can support B2B spare parts search in controlled settings, and document practical design and implementation considerations for hybrid retrieval, structured output enforcement, and performance optimization.