Architecture that splits the generation of results summaries into at least two phases for reduced latency and improved relevance. A first phase generates a summary quickly and thereby enables subsequent modules to begin processing. A second round then executes in parallel to these other modules to offset the latency. The second round can also provide additional contextual information to the summarization module to improve the quality (relevance) of the summaries. Ultimately, the summaries generated in the two phases are merged. The multiple rounds of summarization enable the first round to be cheaper in processing resources to save overall latency, whereas the second phase can be more expensive, since the second phase can be run in parallel with other modules. Additionally, contextual information can be used to build the full content summaries in the second round.