Definition
The initial phase in a multi-stage RAG pipeline where a high-recall subset of potentially relevant documents is extracted from a large-scale corpus, typically using approximate nearest neighbor (ANN) search. It prioritizes retrieval speed and breadth, trading off precision to ensure the 'ground truth' documents are present in the result set before being passed to a computationally expensive reranker.
Initial broad-stroke filtering, not the final re-ordered results or the LLM generation itself.
"A wide-mesh fishing net designed to capture as many relevant fish as possible, accepting that some non-target species will be sorted out later on deck."
- Vector Database(Prerequisite)
- Reranking(Subsequent Step)
- Approximate Nearest Neighbor (ANN)(Component)
- Recall(Primary Metric)
Conceptual Overview
The initial phase in a multi-stage RAG pipeline where a high-recall subset of potentially relevant documents is extracted from a large-scale corpus, typically using approximate nearest neighbor (ANN) search. It prioritizes retrieval speed and breadth, trading off precision to ensure the 'ground truth' documents are present in the result set before being passed to a computationally expensive reranker.
Disambiguation
Initial broad-stroke filtering, not the final re-ordered results or the LLM generation itself.
Visual Analog
A wide-mesh fishing net designed to capture as many relevant fish as possible, accepting that some non-target species will be sorted out later on deck.