Retrieval-augmented generation (RAG) is a method applied in practice to ground large language models in domain knowledge when answering questions. However, it may be difficult for practitioners to build a well-performing RAG solution. To address this challenge, we evaluated RAG system components and various methods across multiple languages, including under-resourced languages of the Baltic States. We compared open-source and commercial embedding models, vector databases, and methods that boost the accuracy of semantic search. Results show minimal performance differences among vector stores, while carefully selected embedding models excel in multilingual settings. We found that indexing condensed text (e.g. thesis statements) reduces the semantic gap between questions and original chunks, thereby increasing retrieval accuracy.