Definition
A retrieval paradigm in RAG that orders candidate documents by the estimated probability of their relevance to a query, typically implemented via the Probability Ranking Principle (PRP). It involves balancing the architectural trade-off between high-latency, high-precision re-rankers (like Cross-Encoders) and low-latency, lower-precision initial retrievers.
Distinguishes between simple geometric vector similarity and the statistical likelihood of a document containing the correct answer.
"A multi-stage filtration system where a wide mesh catches many particles, but a final digital sensor ranks them by their specific purity level before they reach the outlet."
- BM25(Component)
- Cross-Encoder(Component)
- NDCG (Normalized Discounted Cumulative Gain)(Prerequisite)
- Hybrid Search(Component)
Conceptual Overview
A retrieval paradigm in RAG that orders candidate documents by the estimated probability of their relevance to a query, typically implemented via the Probability Ranking Principle (PRP). It involves balancing the architectural trade-off between high-latency, high-precision re-rankers (like Cross-Encoders) and low-latency, lower-precision initial retrievers.
Disambiguation
Distinguishes between simple geometric vector similarity and the statistical likelihood of a document containing the correct answer.
Visual Analog
A multi-stage filtration system where a wide mesh catches many particles, but a final digital sensor ranks them by their specific purity level before they reach the outlet.