Artificial needles to real haystacks: Improving retrieval capabilities in LLMs