1. We pick documents with highest scores from each search result. We are doing this by dropping documents with lowest scores from each search result one by one until it fits into the

max_bytes

window 2. We don’t support sparse vectors right now. 3. Can you give example for rewrite/optimization? 4. We don’t support that unfortunately, but as long as it fits into the window we send all documents in the results as the context for answering the question, so I can’t see any reason for reranking before answering the question. 5. No, we don’t have any chunking method. You should use a library for NLTK for this purpose.