Menu
Use a two-stage retrieval pipeline with reranking to balance recall, precision, and latency.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Use this skill when building, debugging, or improving Retrieval-Augmented Generation systems, including chunking, vector database selection, hybrid search, reranking, multimodal RAG, code documentation RAG, retrieval latency, and production RAG architecture.
Chunk nested documents into parent-child levels so retrieval can move from broad sections to fine-grained passages.
Use semantic boundaries and embedding similarity to chunk text for higher-relevance retrieval.
Route RAG chunking decisions across semantic, hierarchical, sliding-window, contextual-header, and framework-selection strategies.
Use overlapping windows to preserve context across chunk boundaries while controlling retrieval size.
Reduce retrieval latency with caching, batching, and index-level optimization.
| name | multi-pass-retrieval-with-reranking |
| title | Multi-Pass Retrieval with Reranking |
| description | Use a two-stage retrieval pipeline with reranking to balance recall, precision, and latency. |
| allowed-tools | ["Read","Grep","Glob","Bash"] |
| category | retrieval-strategies |
| tags | ["reranking","cross-encoder","two-stage","relevance"] |
Multi-pass retrieval with reranking is a two-stage approach that first retrieves a broad set of candidates using fast bi-encoder search, then refines them using a more accurate but slower cross-encoder reranker. This balances recall and relevance while maintaining acceptable latency.
Single-stage retrieval has inherent trade-offs:
Implement the first-stage retrieval using your vector database.
Why: Bi-encoder retrieval is fast and scalable but produces approximate scores - suitable for the first pass to narrow candidates.
Implement the second-stage reranking for accurate scoring.
Why: Cross-encoders produce more accurate relevance scores by jointly encoding query-document pairs, but are too slow for full-scale retrieval.
Dynamically adjust retrieval parameters based on query characteristics.
Why: Adaptive retrieval ensures complex queries get more candidates for reranking while simple queries remain fast.
Combine all components into a complete retrieval pipeline.
Why: A unified pipeline with timing and explanation makes it easy to optimize and debug the retrieval system.
Add caching and optimizations for production use.
Why: Caching reduces computation for repeated queries, which is common in production systems.
For working examples, see Sentence Transformers CrossEncoder docs, Sentence Transformers rerankers, Sentence Transformers retrieve & re-rank, and the MonoT5 paper.
