Two-Stage Retrieval Architecture
Jump to navigation
Jump to search
A Two-Stage Retrieval Architecture is an information retrieval architecture that combines efficient candidate generation with accurate reranking through sequential processing stages.
- AKA: Retrieve-and-Rerank Architecture, Cascade Retrieval Architecture, Funnel-Based Retrieval Architecture.
- Context:
- It can typically employ First-Stage Retrievers with high-recall objectives.
- It can typically utilize Second-Stage Rerankers with high-precision objectives.
- It can typically combine Bi-Encoder Models with Cross-Encoder Models for efficiency-accuracy tradeoff.
- It can often integrate BM25 Algorithms for lexical matching.
- It can often apply Dense Retrieval Methods for semantic similarity computation.
- It can often incorporate Learning-to-Rank Models for relevance score optimization.
- It can often leverage Reciprocal Rank Fusion for result combination.
- It can range from being a Shallow Two-Stage Retrieval Architecture to being a Deep Two-Stage Retrieval Architecture, depending on its model complexity.
- It can range from being a Homogeneous Two-Stage Retrieval Architecture to being a Heterogeneous Two-Stage Retrieval Architecture, depending on its component diversity.
- It can range from being a Fixed Two-Stage Retrieval Architecture to being a Adaptive Two-Stage Retrieval Architecture, depending on its configuration flexibility.
- It can range from being a Lightweight Two-Stage Retrieval Architecture to being a Heavy Two-Stage Retrieval Architecture, depending on its computational requirement.
- ...
- Examples:
- Legal Two-Stage Retrieval Architectures, such as:
- COLIEE 2025 Retrieval Pipeline, combining BM25 algorithm with legal cross-encoders.
- COLIEE 2024 Hybrid System, using dense retrievers with LLM rerankers.
- Commercial Two-Stage Retrieval Systems, such as:
- ...
- Legal Two-Stage Retrieval Architectures, such as:
- Counter-Examples:
- Single-Stage Retrieval Architecture, which performs retrieval in one pass.
- End-to-End Retrieval Architecture, which jointly optimizes all components.
- Multi-Stage Retrieval Architecture, which uses more than two processing stages.
- See: Bi-Encoder Model, Cross-Encoder Model, Information Retrieval System, Dense Retrieval Method, Reranking Task, BM25 Algorithm, Learning-to-Rank Model, Information Retrieval System Architecture, Neural Information Retrieval Model.