How dow we ensure a proper and deterministic retrieval of citations and references?

We use RAG ( Retrieval-Augmented Generation) on pubmed abstracts and we make sure the pubmed ids returned do exist. There is still the possibility of the LLM giving an answer that doesnt match the content of the retrieved pubmed IDs but it is low. We also evaluate internally against public benchmarks related to literature review. We also know we can improve our daily evaluation tests to analyze the actual answer uses the retrieved pubmed IDs.