Nesyona Research // Data Study

Vector Databases for RAG: Deployment-Shape, License, Hybrid-Search and Pricing Comparison 2026

Cite this dataset: DOI 10.5281/zenodo.20738950 (CC-BY 4.0)

Which vector database fits a RAG workload, given that hybrid search and filtering are now table-stakes? An 11-database comparison by deployment shape, license, and pricing model.

Last updated:
Bottom line: Across 11 vector databases for retrieval-augmented generation, hybrid search and metadata filtering are now near-universal, so the deciding axis is deployment shape (embedded, self-hosted, managed, serverless) matched to your data gravity, not a recall benchmark. The cheapest correct answer is usually the system your data already lives in: pgvector if you are on Postgres, Redis Query Engine if you run Redis. Reach for a dedicated database when you outgrow that: Qdrant for fast self-hosted value, Weaviate for native hybrid, Milvus for billion-scale, Pinecone or turbopuffer for zero-ops.

The comparison matrix

Eleven databases across the four deployment shapes, on the five dimensions that decide fit and cost. yes and no are read from each vendor's own repository, documentation, or pricing page; partial denotes availability through an integration or full-text add-on rather than a single native call. Pricing model is the column to read at scale, not the headline price. The full machine-readable matrix is in data.json.

DatabaseDeployment shapeOpen-source licenseHybrid searchMetadata filteringPricing modelBest for
pgvectorExtension (in Postgres)PostgreSQL Licensepartial (Postgres FTS)yes (SQL)free (pay your Postgres)apps already on Postgres, under ~10M vectors
QdrantSelf-Hosted + ManagedApache-2.0yes (sparse+dense)yes (strong, Rust)usage / free tierfast filtered search at low self-host cost
WeaviateSelf-Hosted + ManagedBSD-3-Clauseyes (native BM25)yesresource-basednative hybrid search and module ecosystem
Milvus / ZillizSelf-Hosted + ManagedApache-2.0yes (2.4+)yesfree OSS / usage (Zilliz)billion-scale distributed workloads
PineconeServerless (Managed)proprietaryyes (sparse-dense)yesusage (serverless)zero-ops fully-managed RAG
ChromaEmbedded + CloudApache-2.0partial (integration)yesfree OSS / cloud usageprototyping and local-first RAG
LanceDBEmbedded + CloudApache-2.0partial (FTS)yesfree OSS / cloudmultimodal and edge / embedded
turbopufferServerless (Managed)proprietaryyes (BM25+vector)yesusage (object-storage)storage-heavy, cost-optimized workloads
VespaSelf-Hosted + ManagedApache-2.0yes (native ranking)yesfree OSS / resource (Cloud)complex ML-driven ranking over text+vectors
MarqoSelf-Hosted + CloudApache-2.0yesyesfree OSS / cloudend-to-end embedding generation + storage
Redis (Query Engine)Self-Hosted + ManagedRSALv2 / SSPLv1 / AGPLv3yes (vector + text)yesfree OSS / Redis Cloudlow-latency vectors where Redis is already in the stack

Deployment shapes, licenses, hybrid support, and pricing models reflect each vendor's public repository and documentation as of June 2026 and change often; verify on the vendor's own page before a purchase decision. Redis is marked partial under open source because version 8 ships a tri-license (RSALv2/SSPLv1/AGPLv3) rather than a single permissive license. The narrative companion to this dataset is Best Vector Databases for RAG 2026.

Methodology

Each cell is read from the vendor's own public repository, documentation, or pricing page as of June 2026. Open-source licenses are verified from the project's GitHub repository (Redis confirmed as the Redis 8 tri-license RSALv2/SSPLv1/AGPLv3, not a single permissive license). Hybrid search is marked yes only where a single native query path is documented; partial denotes availability through an integration or full-text add-on. Metadata filtering is near-universal across the category. Vendor performance and cost claims are attributed, not asserted; no first-party recall or latency benchmark is encoded here.

The dataset tracks five dimensions: deployment_shape, open_source_license, hybrid_search, metadata_filtering, and pricing_model, plus each tool's "best for" fit. The four-shape reference model (Embedded, Self-Hosted, Managed, Serverless) and the rankings were fixed before any monetization check; there is no paid placement and no sponsorship from any database listed.

Open dataset. The full matrix is published at data.json under a CC-BY 4.0 license, free to share and adapt with attribution to Nesyona / Vincent Couey (ORCID).

Save
Dashboard