AI agents · OpenClaw · self-hosting · automation

Quick Answer

Best Hybrid Search Engines for RAG 2026: BM25 + Vectors

Published:

The short answer

Use a hybrid engine whenever your users search for exact strings — names, SKUs, error codes — because pure vector search misses them. As of November 16, 2026:

  1. Qdrant — best for new RAG stacks. Dense + BM25 + RRF in one call, Apache-2.0, generous free cloud tier.
  2. OpenSearch — best full search engine under an open licence. Hybrid query plus normalisation or RRF in a search pipeline.
  3. Weaviate — best managed vector database with hybrid on by default. One alpha parameter weights keyword vs vector.
  4. Elasticsearch — best if you already pay for Elastic Enterprise; RRF is not on the free tier.
  5. Meilisearch and Typesense — best for app and site search where hybrid is a feature, not a project.

The comparison

EngineLicence / versionHow it fusesPrice (Nov 16, 2026)Best for
QdrantApache-2.0, v1.20.1 (Nov 13, 2026)Query API: prefetch dense + sparse/BM25 (server-side Qdrant/bm25), fuse with RRF or DBSF; Formula Query for custom scoringFree cloud: 0.5 vCPU, 1 GB RAM, 4 GB disk; Standard usage-basedRAG backends, filtering-heavy retrieval
OpenSearchApache-2.0, 3.9.0 (Sep 29, 2026)hybrid query + normalisation processor (score-based) or score ranker processor (RRF, since 2.19)Free self-host; managed via cloud providersTeams needing full-text features, logs and RAG in one engine
Weaviatev1.40.5 (Nov 10, 2026)BM25F + vector, alpha weight, relativeScoreFusion or ranked fusionFree: 100,000 objects, 1 GB memory; Flex from $45/month; Premium contractManaged vector DB with built-in embeddings
ElasticsearchElastic License 2.0 / AGPL, v9.5.6 (Nov 2, 2026)RRF retriever and linear retrieverRRF requires Enterprise subscriptionExisting Elastic Enterprise customers
MeilisearchMIT core + BUSL-1.1 Enterprise Edition, v1.56.2 (Nov 9, 2026)Built-in hybrid with semantic ratioCloud from $20/month (fixed per instance); self-host freeSite and app search with typo tolerance
TypesenseGPL-3.0, v30.2vector_query with alpha; rank_fusion_score; optional rerank_hybrid_matchesSelf-host free; Typesense CloudLow-latency app search

1. Qdrant — best for new RAG stacks

Qdrant’s Query API makes hybrid a single request: two prefetch branches (a dense embedding and a BM25 vector that Qdrant computes on the server from the raw text) and a FusionQuery with RRF or DBSF on top. No sparse-embedding client library is needed. Its payload filtering is strong, which matters for multi-tenant RAG. The free cloud tier (single node, 0.5 vCPU, 1 GB RAM, 4 GB disk) is enough to prototype; production clusters are priced by resources, with Terraform and Pulumi automation on every tier.

2. OpenSearch — open-source search engine with real hybrid

OpenSearch has had the hybrid compound query since 2.11. You attach a search pipeline that combines the clauses either with the normalisation processor (rescales scores, keeps margins between documents) or with the score ranker processor (RRF, introduced in 2.19). Everything is Apache-2.0, so RRF costs nothing — the key difference from Elasticsearch. Choose it when you also need aggregations, analysers for many languages and mature operations tooling.

3. Weaviate — managed hybrid by default

Weaviate runs BM25F and vector search together and fuses them; the alpha parameter (0 = keyword only, 1 = vector only) is the only knob most teams touch, and relativeScoreFusion is the default fusion. The free cloud tier holds 100,000 objects with 1 GB memory and 10 GB disk plus 2,000 embedding requests a day; Flex starts at $45 a month pay-as-you-go.

4. Elasticsearch — only if you are on Enterprise

Elasticsearch’s retriever API is generally available and the RRF retriever combines BM25 and kNN cleanly — but Elastic’s subscription matrix puts RRF and the linear, rule and re-ranker retrievers in the Enterprise tier only. On Basic you can still run both searches and merge the lists in application code.

Both are built for instant, typo-tolerant search boxes and added vector search later. Meilisearch Cloud is a fixed monthly price per instance from $20, regardless of query volume; note that parts of the code are under the Business Source Licence. Typesense (GPL-3.0) exposes the fusion weight directly through alpha in vector_query.

How to choose

  • Building a RAG backend from scratch: Qdrant.
  • Need full-text search features and an open licence: OpenSearch.
  • Want a managed database that also embeds for you: Weaviate.
  • Already on Elastic Enterprise: stay on Elasticsearch.
  • Search box on a website or app: Meilisearch or Typesense.
  • Then add a reranker on the top 50–100 fused results — it usually improves answers more than tuning fusion weights. See best reranker models and how to add a reranker to a RAG pipeline.

Last verified: November 16, 2026.

Sources