Retrieval-augmented generation (RAG) and long-context language models (LCLMs) address the contextual limitations of LLMs, but determining the optimal external context to retrieve remains an unresolved challenge. Adaptive-k retrieval is a simple and effective single-pass method that adaptively selects the number of phrases based on the distribution of similarity scores between the query and candidate phrases. It does not require model fine-tuning, additional LLM inference, or modifications to the existing search-reader pipeline.
