hybrid retrieval

**Hybrid retrieval** is the **search strategy that combines dense semantic retrieval and sparse lexical retrieval to improve overall relevance** - it leverages complementary strengths of both paradigms. **What Is Hybrid retrieval?** - **Definition**: Retrieval pipeline that merges rankings or scores from dense and sparse retrievers. - **Fusion Methods**: Weighted score combination, reciprocal rank fusion, or learned rank aggregation. - **Coverage Benefit**: Dense handles semantic similarity while sparse preserves exact-term matches. - **System Requirement**: Needs calibrated scoring and deduplication across candidate lists. **Why Hybrid retrieval Matters** - **Recall and Precision Balance**: Improves broad relevance without sacrificing keyword accuracy. - **Robustness**: Performs better across heterogeneous query types than single-mode retrievers. - **Enterprise Fit**: Handles both natural-language questions and structured identifier lookups. - **RAG Quality Gain**: Better retrieval quality directly improves generation factuality. - **Failure Mitigation**: Reduces missed documents from semantic-only or lexical-only blind spots. **How It Is Used in Practice** - **Dual Retrieval Stage**: Run dense and sparse search in parallel over same corpus. - **Fusion Calibration**: Tune blend weights using offline relevance benchmarks. - **Re-ranking Layer**: Apply cross-encoder ranking on fused candidates for final precision. Hybrid retrieval is **a high-performing default architecture for production search and RAG** - combining semantic and lexical signals yields stronger, more consistent retrieval quality across real workloads.