Improvement on vector documents summarizations

[KartnDev]

Currently we do 2 steps:

1. Find more relevant documents belongs to query
2. Rerank top3

Instead:

1. Find more relevant documents belongs to query
2. Rerank top3
3. Summarize by embedded llm into single structured & compressed doc

OR even

1. Find more relevant documents belongs to query
2. Summarize by embedded llm into single structured & compressed doc (more accurate)

BTW this issue raising following critical issue with agents (and indeed self chat also):