Keyword search finds the exact words. Vector search finds the meaning. Hybrid search runs both and merges the results.
How it works
The query goes to two retrievers at once:
- Keyword: BM25 over an inverted index. Strong on names, codes and rare terms.
- Vector: nearest neighbours over embeddings in a vector database. Strong on paraphrases and synonyms.
The two scores are on different scales, so most systems merge by rank or by normalized score.
Reciprocal rank fusion (RRF), from Cormack, Clarke and Büttcher (SIGIR 2009), adds up a term for each list a document appears in. Elasticsearch documents the loop as score += 1.0 / (k + rank(result(q), d)), with the rank constant k defaulting to 60. Its docs say “RRF requires no tuning.”
A worked example with k = 60:
| Document | BM25 rank | Vector rank | RRF score |
|---|---|---|---|
| A | 1 | 3 | 1/61 + 1/63 = 0.0323 |
| B | 2 | 1 | 1/62 + 1/61 = 0.0325 |
| C | not found | 2 | 1/62 = 0.0161 |
B wins because both retrievers rank it high.
Weighted fusion is the other common method. Weaviate’s default since v1.24, relativeScoreFusion, normalizes each search’s scores and combines them as a scaled sum. Its alpha setting weights the two: 0 is keyword only, 1 is vector only, and 0.75 is the server default. The docs recommend setting alpha explicitly: the 0.75 default applies only when a request reaches the server with no alpha, and not every client library leaves it unset, so the real weighting can differ.
After fusion, many pipelines add a reranker over the top results.
Why it matters
- Recall. Real queries mix exact terms and loose wording. One retriever alone misses part of them.
- Arabic. Arabic spelling varies (alef forms, ta marbuta, diacritics), which hurts exact keyword matching unless text is normalized. Adding vector search gives a second path to the right chunk. See Arabic RAG search and Arabic text normalization for search.
- Fewer tokens. Better retrieval means fewer chunks sent to the LLM per question.
On BroadNet Inference
Hybrid search runs in your search stack: Elasticsearch, OpenSearch, Weaviate or similar. BroadNet does not offer embeddings or rerank endpoints, so produce embeddings with another provider or a local model. Send the merged top chunks to a BroadNet chat model such as Qwen3.8 27B or GLM 5.3 over the Chat Completions API. For small corpora, compare against putting the documents in a long context; see long context vs RAG.