Using RAG improves search efficiency by selecting only the results that seem most relevant for computation. Although embeddings and rerankers can perform quite well, general or abstract information often fails to produce meaningful answers when the data is heavily chunked and lacks sufficient context.
You can increase the number of retrieved results, but this may lead to hitting the model’s context or token limit.
The following article may help: results can often be improved by adding more contextual information into each vector during the chunking process.