Retrieval-Augmented Generation for Sustainable Material Data Handling in Automotive Value Chain

Sammanfattning

Applying large language models (LLMs) to industrial material data workflows has thepotential to improve efficiency. However, conventional LLMs are limited by hallucinations,depend on proprietary training data, and are costly to update. This thesis exploresRetrieval-Augmented Generation (RAG) as an alternative approach, in which an LLMgenerates responses grounded in a restricted, domain-specific corpus of documents anddatabases and provides source citations for them. The study is carried out in an industrialsetting with two main data domains: an internal SQL materials database with tabularmaterial properties, and a corpus of unstructured textual documents, including supplierdocuments, corporate standards, and environmental product declarations. A RAG systemis developed that (1) indexes both textual and tabular data, (2) retrieves relevant chunksvia dense vector search, and (3) generates source-grounded responses.The work investigates whether a RAG model that explicitly integrates both domains canoutperform a baseline tuned for unstructured text and explores which tabular serializationformat yields the most semantically informative embeddings for pretrained embeddingmodels. To achieve this, we first constructed LLM-based pipelines to generate documentandtable-based test sets with ground-truth chunk annotations, and implemented a modularRAG pipeline with separate indices for textual and tabular data. Then, we experimentedwith multiple retrieval strategies, ranging from concatenating the retrieval results tousing cross-encoders to weigh them. In addition, several fusion strategies were tested toevaluate whether they could improve retrieval accuracy when operating across differentdomains. Experiments are conducted comparing nine tabular serialization strategies,studying performance as a function of index size, chunk size, and top-k, and evaluatingdifferent fusion modes and embedding models. The evaluation metrics used are Hit Rate,Recall, Precision, F1-score, and Mean Reciprocal Rank.Results show that enriched serialization, which converts tabular rows into natural-languagestatements, yields stronger tabular retrieval performance than a standard key-value-basedformat, without degrading performance on document retrieval. Larger chunk sizes andhigher top-k values systematically improve retrieval metrics, highlighting both the difficultyof relying solely on similarity search and the benefits of cross-encoder reranking on largercandidate sets. A domain-aware weighted fusion retriever further improves overall retrievalperformance over the optimized baseline with only moderate computational overhead. Thesefindings demonstrate that semantically rich tabular representations and domain-awarefusion can enhance RAG performance on heterogeneous industrial material data.