Explain the internal working of tree_summarize response synthesis?

tree_summarize is built for queries that genuinely need to draw on many retrieved Nodes holistically, such as "summarize the key risks across all these reports," where sequentially refining through Nodes one by one, as refine does, tends to over-weight whichever Node happened to be processed first or last.

Instead, it works bottom-up like a tree. Retrieved Nodes are grouped into batches sized to fit the LLM's context window, and each batch is summarized independently and in parallel. If that produces more than one intermediate summary, those summaries are themselves treated as a new set of "nodes" and the same batching-and-summarizing step repeats recursively, forming a tree of summaries collapsing upward. This continues until only a single summary remains, which becomes the final answer.

The trade-off is more LLM calls than compact when there are many Nodes, since each level of the tree is its own round of calls, but it avoids the early-Node bias of refine and tends to produce a more balanced synthesis across a large or diverse set of retrieved content.