Vector databases in 2026: Pinecone vs Weaviate vs Qdrant vs Milvus vs Chroma

A retrieval application needs more than somewhere to store embeddings. It must combine relevance with filters, updates, access boundaries, and an operating model the team can support. A prototype that finds a similar paragraph on a laptop does not yet answer how production searches behave while data changes or one customer must remain isolated from another.

Pinecone, Weaviate, Qdrant, Milvus, and Chroma all support vector-oriented retrieval, but their deployment and commercial models differ. The useful comparison is between complete retrieval systems for a particular workload, rather than vector-count headlines or unrelated vendor benchmarks. Start with the data and queries your application actually needs to serve.

What this decision involves

Vector retrieval finds items represented by numerical embeddings. The application may also need exact terms, metadata constraints, or a reranking stage. A useful database choice preserves those requirements without forcing the application to download a large candidate set and repair every result afterward.

This differs from choosing a general-purpose relational host. Our managed PostgreSQL comparison covers that broader database-operations decision. A dedicated retrieval layer may coexist with a relational source of truth. Decide which system owns documents, permissions, and deletion before choosing how copies and embeddings reach the search index.

The quick difference

Product	Best fit	Main strength	Watch for
Pinecone	Managed retrieval operations	Hosted indexes and usage-oriented service plans	Plan minima, region controls and optional services
Weaviate	Semantic plus keyword workflows	Objects, vectors and hybrid search	Cloud tier, embedding and agent charges
Qdrant	Filtered vector applications	Payload indexes and deployment choices	Index design, cluster sizing and cloud limits
Milvus	Teams choosing deployment architecture	Lite, Standalone and Distributed options	Self-operated infrastructure versus Zilliz Cloud
Chroma	Application-led retrieval development	Documents, embeddings and metadata queries	Local/server/cloud distinctions and usage meters

Pinecone

Pinecone provides managed vector database infrastructure, with current pricing covering dense, sparse, and full-text indexes and related inference or assistant services. It is relevant when the team wants to focus on retrieval behavior while the provider operates the database service.

The current site separates Starter, a flat-price Builder plan, Standard with minimum monthly usage, and Enterprise with a higher minimum and additional controls. Standard is presented with a $50 monthly minimum, after which usage determines spending. Enterprise adds capabilities such as private endpoints and bring-your-own-cloud. These are different purchasing arrangements, not interchangeable prices for the same unrestricted deployment.

The strength is managed operations. The tradeoff is understanding the service’s meters and plan-specific boundaries. Use the selected currency and region in a quote, then estimate ingestion, storage, search, inference, and support separately. In a pilot, update and delete representative documents, constrain a query by tenant, and verify what the application sees before and after those changes become searchable.

Weaviate

Weaviate is an open-source database that stores objects and their vector representations, with semantic and hybrid search capabilities. Its documentation also describes supporting retrieval-augmented applications and connected services. It suits teams that want the relationship between object data and retrieval to remain visible in the database model.

The current cloud offer uses Free, Flex, and Premium. Free has a limited cluster, object count, memory, disk, collection, and tenant allowance. Flex is a shared-cloud, monthly pay-as-you-go offer starting at a displayed $45 per month; Premium uses a prepaid contract and offers shared or dedicated deployment. Embeddings and Query Agent have their own allowances or usage charges.

The strength is combining semantic and keyword-oriented retrieval with structured objects. The tradeoff is selecting a data model and operational tier that fit the workload. Compare a keyword-heavy query, a conceptual query, and a tenant-filtered query using the same reviewed documents. Confirm that any connected model or agent service is available in the environment you intend to deploy.

Qdrant

Qdrant emphasizes vector search with payload filtering. Its documentation explains payload indexes and how they interact with search, making metadata design part of the retrieval architecture. That matters when a result must satisfy conditions such as customer, region, availability, or document type as well as semantic similarity.

Qdrant can be self-operated or purchased through managed offerings. Its cloud pricing currently distinguishes a small single-node Free tier, Standard with dedicated resources and usage-based pricing, and Premium with minimum spending and additional controls. Hybrid Cloud runs managed clusters on the customer’s infrastructure; Private Cloud addresses more isolated deployment needs. These choices carry different infrastructure and support responsibilities.

The strength is explicit control over filtered retrieval. The tradeoff is planning indexes and resource use. Cloud strict-mode limits can reject filtering on unindexed fields, so test the intended schema early. During evaluation, create the metadata indexes first, run a restrictive filter, and inspect behavior as the data distribution changes. Size the cluster from that workload rather than an unfiltered demonstration.

Milvus

Milvus provides several deployment paths under a shared client-side API. Official documentation distinguishes Milvus Lite for local application use, Standalone for a single-machine server, and Distributed on Kubernetes. This gives teams a way to discuss architecture separately from the retrieval code they begin with.

The strength is choosing an operating shape appropriate to the application. The tradeoff is that running the open-source software means owning the surrounding infrastructure, upgrades, recovery, and capacity planning. A compatible API does not make every deployment operationally equivalent or guarantee that all features have identical boundaries.

Zilliz Cloud is a separately purchased managed service associated with Milvus. Its pricing distinguishes free, serverless, dedicated, and enterprise-oriented arrangements; do not describe a Zilliz allowance as a Milvus software license fee. In a pilot, document the exact deployment and version, then check import, deletion, backups, and a move from the development environment. Include the cost of storage, compute, support, and the people operating the system.

Chroma

Chroma combines document and metadata storage with embedding-oriented retrieval. Current documentation describes dense, sparse, hybrid, full-text, and metadata search, with choices for running Chroma and connecting clients. It is relevant when application developers want a direct route from stored content to a useful retrieval workflow.

Keep local or self-operated Chroma separate from Chroma Cloud. The cloud pricing page meters writing, stored data, queried data, and returned network data. Starter is shown as a zero monthly base plus usage, while Team has a monthly base plus usage and included credits; enterprise arrangements are custom. A zero base price is not unlimited free retrieval.

The strength is an application-friendly retrieval model. The tradeoff is validating the deployment and feature combination the application will actually use. Start with your real embedding dimensions, document sizes, and metadata conditions. Then check how updates, collection boundaries, exports, and concurrent client requests behave. Estimate cloud spending from the size of data read and returned, not only the number of API calls.

How to choose

Create a small evaluation collection with reviewed relevance judgments. Include easy semantic matches, exact identifiers, ambiguous queries, and results that must be excluded by permissions. Keep the embedding model, chunking strategy, and candidate set consistent when comparing databases. Otherwise a different preprocessing choice can be mistaken for a database advantage.

Measure the workflow you will operate: ingestion, updates, deletes, filtered retrieval, and recovery. Record the deployment version and search settings alongside any results. A fast unrestricted query is not evidence that a restrictive tenant query will behave the same way. Likewise, a vendor benchmark does not replace your own acceptance criteria.

Compare total costs in explicit units. Include model inference, index storage, replicas, query and write meters, networking, support, and engineering time. For dollar-denominated figures, confirm the contractual currency and taxes rather than assuming every regional invoice matches the public display. Finally, plan export and reindexing. Changing an embedding model or leaving a provider can require reading the original source data again, so preserve a clear source of truth.

Selection checklist

Keep embeddings, chunking, query sets, and relevance judgments consistent during evaluation.
Test metadata filters and tenant exclusions together with ordinary semantic queries.
Demonstrate updates, deletion, backup recovery, and export on the intended deployment.
Model storage, reads, writes, inference, network traffic, support, and operational labor.
Confirm region, version, paid controls, quotas, and a practical reindexing plan.

Conclusion

Pinecone suits a managed-service preference, Weaviate combines objects with semantic and hybrid search, Qdrant makes filtered retrieval a central design concern, Milvus offers distinct deployment architectures, and Chroma provides an application-oriented retrieval workflow. Choose according to your data boundaries and operating capacity. A successful shortlist preserves relevance, filtering, and recovery requirements at a cost the team can explain and sustain.

References

Official product, pricing, and documentation sources reviewed on October 15, 2026. Suggested evaluation tasks are buying criteria, not claims of hands-on testing.