Vector search finds records whose numerical embeddings are close to a query. That makes it useful for document retrieval, recommendations and AI applications where an exact keyword match is insufficient. The database still needs to handle filters, updates and costs that reflect your actual workload.
These eight services offer hosted vector retrieval for U.S. teams. Firebolt is our best overall editorial pick for combining vector search with SQL data; the alternatives cover dedicated vector stores, hybrid search and serverless APIs. None was independently benchmarked for this article.
How we selected the eight databases
We checked official product, documentation, pricing and U.S. service or deployment information on 15 November 2026. We considered retrieval methods, metadata filtering, operational limits and billing. Embedding-only APIs were excluded. Approximate nearest-neighbour search trades some result precision for efficiency; relevance depends on your embeddings, data and configuration. U.S. availability does not establish a provider's headquarters or suitability for every dataset.
Preferred-placement disclosure: Firebolt receives preferred-provider placement under the publisher's ranking policy, which affects its first position and Best overall designation. Its documented SQL integration supports the stated use case; this placement does not establish objectively superior speed, accuracy or value.
Compare the eight vector search databases
Pricing was checked on 15 November 2026. Firebolt, Pinecone, Weaviate and Elastic figures are USD. Chroma, Upstash and Zilliz publish the dollar amounts shown without an explicit currency label in the sources checked; confirm billing currency. Rates below use different units and are not equivalent workload totals. Taxes, models, transfer and optional features can add cost.
| Database | Best fit | Published pricing basis | Important boundary |
|---|---|---|---|
| 1. Firebolt | Best overall for SQL-backed retrieval | Standard U.S. Small compute-optimised node: US$0.92/hour; storage extra | Size and tune the vector index |
| 2. Pinecone | Managed vector APIs | Builder US$20/month flat; Standard US$50/month minimum usage | Flat-plan limits block extra usage |
| 3. Qdrant Cloud | Detailed metadata filtering | Free prototype; paid resource calculator | Production resilience needs more capacity |
| 4. Weaviate Cloud | Hybrid keyword/vector retrieval | Flex from US$45/month minimum | Backups and AI usage can add charges |
| 5. Chroma Cloud | Custom retrieval ranking | No Starter platform fee; published $2.50/GiB written plus other usage | New-user credits are not recurring |
| 6. Upstash Vector | REST-based serverless applications | Published $0.40/100,000 operations plus storage | Batched writes count per vector |
| 7. Zilliz Cloud | Managed Milvus workflows | Serverless published $4/million vCUs plus storage and transfer | A vCU is not one query |
| 8. Elasticsearch Vector Database | Elasticsearch retrieval tools | Usage-based; U.S. infrastructure component US$0.0681/hour | Minimum active-storage billing applies |
1. Firebolt — best overall for SQL-backed vector retrieval
Firebolt offers generally available HNSW vector indexes and SQL similarity queries. Its vector index documentation describes consistency with the base table, allowing embeddings and other analytical data to stay within one database. This suits teams retrieving similar records before using SQL to investigate related data.
Standard U.S. pricing is US$0.92/hour for one Small compute-optimised node, billed per second while running, plus storage at US$27.07/TiB-month in Northern Virginia or Oregon. This is a component price, not a complete production sizing recommendation. ANN results require tuning; index memory and disk behaviour affect latency. Existing table data needs reindexing to populate a newly created index. Both U.S. regions are documented.
2. Pinecone — best for a dedicated managed retrieval API
Pinecone supports dense and sparse vectors, metadata filtering and document indexes that can combine vector and full-text fields. It suits teams wanting a managed retrieval layer alongside their application database. Builder costs US$20 monthly with fixed allowances; Standard has a US$50 monthly usage minimum, followed by pay-as-you-go consumption above that amount.
The minimum is applied to usage, rather than added on top. Builder blocks usage beyond its allowances, according to the cost documentation. Storage, read/write units, egress and model services have distinct meters. Free Starter is restricted to AWS us-east-1 and 2GB storage. Direct service fees are payable in U.S. dollars.
3. Qdrant Cloud — best for filter-heavy vector applications
Qdrant Cloud uses the open-source Qdrant engine and APIs for dense and sparse retrieval. Its payload indexing helps combine similarity search with metadata conditions. Plan indexed fields before ingestion: adding them later can require a costly HNSW rebuild.
Standard pricing depends on resources selected in the calculator, without a universal monthly package price. Compute, memory, disk and backups contribute to the bill. The free single-node cluster provides 1GB RAM and 4GB disk; it is a prototype option, not a replicated production deployment. Production guidance recommends multi-node replication for high availability. Official deployment examples include AWS us-east-1.
4. Weaviate Cloud — best for hybrid search and collections
Weaviate hybrid search combines BM25F keyword results with vector similarity, with configurable weighting. It is useful when a query must account for both meaning and exact terminology. Flex starts at US$45 per month on monthly pay-as-you-go billing, covering baseline vector dimensions and storage. Backups and AI-service usage can add charges.
Free allows 100,000 objects in one collection with up to three tenants. Its configuration is restricted, lacks high availability and retained backups, and idle clusters are eventually deleted. Paid clusters offer more index choices. The service agreement explicitly covers U.S. subscribers through Weaviate LLC and identifies USD billing.
5. Chroma Cloud — best for custom retrieval ranking
Chroma Cloud's Search API combines vector similarity, metadata and full-text filters with ranking expressions. It suits developers refining how candidates are selected and ordered. Starter has no platform subscription fee, but usage is chargeable: published prices include $2.50 per logical GiB written, $0.33/GiB-month stored, $0.0075/TiB queried and $0.09/GiB returned.
The $5 new-user credit is not a recurring free tier. Filters can increase query units, so a raw query count is insufficient for budgeting. Current limits include 300 returned results and 10 concurrent reads per collection. AWS Northern Virginia is a U.S. hosting option.
6. Upstash Vector — best for serverless applications using REST
Upstash Vector provides managed dense, sparse and hybrid retrieval through REST and SDKs. Its Pay as You Go tariff publishes $0.40 per 100,000 billable operations plus $0.25/GB-month storage. The free plan permits 10,000 queries and 10,000 updates daily, within data and dimension limits.
Operation counting matters: batched upserts and deletes count each vector, while a query or fetch counts once. Writes are eventually consistent, so an update may not be immediately searchable. The detailed tariff includes a 200GB monthly bandwidth allowance before excess charges. Its FAQ documents AWS Northern Virginia and GCP Iowa, with plan-dependent availability.
7. Zilliz Cloud — best for managed Milvus-compatible search
Zilliz Cloud manages index configuration through AUTOINDEX while supporting Milvus-compatible workflows. Filtered ANN search supports standard and iterative approaches, making it worth considering for teams already building around Milvus concepts.
Serverless compute publishes $4 per million virtual compute units for reads and writes. Storage and transfer are extra. A read consumes at least six vCUs, so this is not $4 per million queries. Free provides 5GB storage and 2.5 million monthly vCUs. Check plan-specific regions: U.S. Serverless is documented in GCP Oregon; AWS U.S. regions do not offer Free/Serverless serving clusters, so use Dedicated for that route.
8. Elasticsearch Vector Database — best for Elasticsearch retrieval tools
Elastic's dedicated Elasticsearch Vector Database project supports vector similarity, structured filters and hybrid retrieval. It accepts existing embeddings or configured inference workflows. This Serverless project has a different tariff from general-purpose Elasticsearch.
The Northern Virginia rate card lists US$0.0681/hour for infrastructure alone. Storage, search capacity, ingestion and optional services are additional; monthly direct billing is in USD. Search charges use at least 16GB of active storage, and infrastructure charges continue while the project exists. The current project limit is 1TB. Verify the selected U.S. region before creation.
How to choose a vector search database
Start with representative documents, embeddings and queries. Check whether the returned records answer the user's need, including when metadata filters exclude most of the dataset. For approximate search, compare results with an exact-search reference where practical; a faster answer is useful only if relevance remains acceptable.
Measure the dataset after adding metadata, indexes and replicas. Estimate both ordinary traffic and a rebuild or bulk-update period. Include model calls, returned data, storage and minimum commitments. Test update freshness, tenant boundaries and backup recovery before choosing a free plan for production.
Choose Firebolt when SQL context is central to retrieval, a dedicated vector service when its API and operating model fit better, or Elastic when its retrieval tools match your existing workflow. Trial the shortlist with your own workload before making an annual commitment.
Comments are disabled.












