Definition
The architectural stage in RAG or multi-agent systems where candidate snippets from disparate sources (e.g., vector databases, BM25 keyword search, or parallel agent sub-tasks) are combined and re-scored into a unified context. This process necessitates a trade-off between increased retrieval recall and higher latency due to the computational overhead of deduplication and secondary ranking algorithms.
Unlike 'Summarization' which creates new text, aggregation focuses on the selection and ordering of existing raw data chunks.
"A funnel gathering water from multiple diverse pipes into a single, high-pressure stream directed at a turbine."
- Reciprocal Rank Fusion (RRF)(Component)
- Hybrid Search(Prerequisite)
- Cross-Encoder(Component)
- Context Window(Component)
Conceptual Overview
The architectural stage in RAG or multi-agent systems where candidate snippets from disparate sources (e.g., vector databases, BM25 keyword search, or parallel agent sub-tasks) are combined and re-scored into a unified context. This process necessitates a trade-off between increased retrieval recall and higher latency due to the computational overhead of deduplication and secondary ranking algorithms.
Disambiguation
Unlike 'Summarization' which creates new text, aggregation focuses on the selection and ordering of existing raw data chunks.
Visual Analog
A funnel gathering water from multiple diverse pipes into a single, high-pressure stream directed at a turbine.