BM25F (which is based on IDF) to quickly identify relevant past responses based on the question and reply they got. At each summarization step you can go throu

BM25F (which is based on IDF) to quickly identify relevant past responses based on the question and reply they got. At each summarization step you can go through and find which Q/A pairs have the highest relevance across all of the other questions in the current batch, which helps solve LLMs tendencies to get stuck in high similar repetitions