Contextra executes a **Hybrid Retrieval Pipeline** engineered for maximum precision and recall: 1. **Dual Retrieval**: - **Vector Search**: Performs cosine similarity search over vector embeddings in Qdrant (using HNSW index). - **Keyword Search**: Performs BM25 full-text match over chunk text in PostgreSQL / Redis payload index. 2. **Reciprocal Rank Fusion (RRF)**: Scores from both retrievers are merged using the RRF formula: $$\text{RRF\_Score}(d) = \sum_{m \in M} \frac{1}{k + r_m(d)}$$ where $k=60$ and $r_m(d)$ is the rank of document $d$ in system $m$. 3. **Reranking & Context Assembly**: The top 25 merged candidates are passed to a cross-encoder Reranker model, returning the final Top K chunks bounded by your token budget. ```rust // libs/retrieval/src/hybrid.rs pub async fn fuse_rankings( vector_results: &[ScoredChunk], keyword_results: &[ScoredChunk], k: f64, ) -> Vec<ScoredChunk> { // RRF implementation in Rust } ```