# Hybrid Retrieval

Combines semantic (dense vector) and keyword (sparse/BM25) search so you catch both meaning and exact terms.

## Why the CCAR-P exam tests this

It is the go-to remedy when pure vector search fails on names, codes, and jargon, and P3 tests whether you know why to blend the two.

## Definition

**Hybrid retrieval** runs **dense** semantic search (embeddings, good at paraphrase and conceptual match) alongside **sparse** lexical search (**BM25**/keyword, good at exact tokens — product SKUs, error codes, acronyms, names) and fuses the results. A common fusion method is **Reciprocal Rank Fusion (RRF)**, which merges ranked lists without needing comparable raw scores.

Each mode covers the other's blind spot. Pure vector search can miss an exact identifier that shares no semantic neighborhood; pure keyword search misses synonyms and intent. Blending them **raises recall** across query types, which is why hybrid is a strong default for enterprise corpora full of domain-specific terms.

Hybrid is a **first-stage** technique that produces a better candidate set; it pairs naturally with **reranking** as the precision second stage. Boosting recall first and refining order second is the canonical high-quality retrieval stack.

## Exam trap

Hybrid retrieval improves recall of the candidate set — it does not by itself put the single best chunk on top; that ordering job belongs to reranking.

## Commonly confused with

- **reranking** — Hybrid = broaden/recall across dense+sparse; reranking = reorder/precision. Complementary, not substitutes.
- **rag-pipeline** — Hybrid retrieval is a retrieval technique used inside a RAG pipeline, not the full pipeline.

## Related

- rag-pipeline
- reranking
- chunking-strategy

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P3). Not affiliated with Anthropic.