if i understand the problem correctly - you'd like to feed xMM documents directly into an LLM so that it uses this context to "reason" answers to questions, vs

if i understand the problem correctly - you'd like to feed xMM documents directly into an LLM so that it uses this context to "reason" answers to questions, vs offload the retrieval to a vector db and merely assemble results into an "answer"? and since your dataset is large, the longest context windows are insufficient.