Scope & boundaries
This guide covers the embedding stores and retrieval indexes that AI applications query, evaluated on recall, filtering, scale and operational cost.
It does not cover the model layer and the platform around it (Generative AI & LLM Platforms), or the end-to-end search and retrieval application (Enterprise Search & Agentic RAG Platforms).
Executive Summary
Check whether the database you already run supports vector search before buying a separate one, and decide early who will operate it.
A vector database stores embeddings, which are numeric representations of text, images or other content, and returns the entries closest to a query. Two kinds of product do this job. Dedicated vector databases are built for it. The other kind adds vector search to a relational database, document database, search engine or data platform that many teams already run.
Most of the products in this guide can be self-hosted. Six of the eight vendors named here document a way to run their software on servers or Kubernetes clusters you operate. The other two sell vector search as a service; one of them can also deploy into your own cloud account. Section 3 sets out what each vendor’s own documentation says.
Vector search is no longer sold only by vector database companies. Database and cloud vendors have added vector types, indexes and storage to existing products, and some vector search products have changed owners or names. Section 2 records those changes from the vendors’ own pages. This guide gives no market-size figure.
The Vector Database Market: Who Sells Vector Search Now
The vector database market now includes the large database and cloud vendors. Oracle states that Oracle AI Vector Search brings AI-powered similarity search to business data without managing and integrating multiple databases. Oracle states that Oracle AI Vector Search capabilities include document load, transformation, chunking, embedding, similarity search, and RAG with LLMs, available natively or through APIs within the database. Microsoft states that the vector data type and vector functions are generally available in SQL Server 2025, Azure SQL Database, Azure SQL Managed Instance, and SQL database in Fabric. Microsoft states that vector indexes are preview features in SQL Server 2025. AWS states that Amazon S3 Vectors delivers purpose-built, cost-optimized vector storage for AI agents, inference, RAG, and semantic search. AWS states that S3 Vectors delivers subsecond latency for infrequent queries and as low as 100 milliseconds for more frequent queries.
Ownership has changed as well. MongoDB announced on February 24, 2025 that it has acquired Voyage AI, which makes embedding and reranking models. A request for datastax.com lands on ibm.com. IBM states that Astra DB and Hyper-converged Database (HCD) enhance the NoSQL database of IBM watsonx.data with vector capabilities. IBM states that Hyper-converged Database is the solution for organizations running their database resources on-premise or via private cloud. Chroma states that Chroma is joining Replit. Chroma describes itself as infrastructure supporting vector, full-text, regex, and metadata search, built on object storage, and open-source under Apache 2.0.
For a buyer, the result is more choice and more overlap. The same workload can often run in a dedicated vector database, in a vector index inside the operational database, or in a search engine. The difference shows up in operations, in filtering and hybrid search, and in cost at your vector count.
Licenses differ by product and have changed before. Elastic states that in September 2024 it was adding the OSI approved AGPLv3 license as an option alongside SSPL and the Elastic license. Check the license of the exact version and distribution you would deploy.
Can You Self-Host a Vector Database?
Yes, for most products in this guide. Self-hosting means you run the database on your own servers, virtual machines or Kubernetes clusters, and your team handles upgrades, backups, security and capacity. Vendors also sell two arrangements in between: a managed service deployed into your own cloud account, and a hybrid model in which the vendor’s control plane manages clusters on your infrastructure.
The open-source vector databases document self-managed deployment directly. Weaviate states that it is available as a hosted service, Weaviate Cloud (WCD), or as a self managed instance, and that self-managed instances use the same Weaviate Database as WCD. Weaviate lists Docker for local evaluation and development, and Kubernetes from development to production. Milvus offers three deployment modes: Milvus Lite, a Python library; Milvus Standalone, a single-machine server with all components bundled into a single Docker image; and Milvus Distributed, deployed on Kubernetes clusters. Qdrant states that Qdrant Private Cloud lets you manage Qdrant database clusters in any Kubernetes cluster on any infrastructure, without any connection to the Qdrant Cloud Management Console. Qdrant states that with Qdrant Hybrid Cloud you create and manage database clusters from the Qdrant Cloud UI while all Qdrant databases operate solely within your network, using your storage and compute resources.
Vector search inside an existing database runs wherever that database runs. pgvector is an extension you compile and install, and it supports Postgres 13 and later. Elastic lists Elastic Cloud Hosted, Elastic Cloud Serverless, self-managed deployment on your own premises, and Elastic Cloud on Kubernetes among its deployment options. MongoDB states that Search and Vector Search are generally available in Enterprise Advanced and Community Edition. MongoDB states that if you run MongoDB Vector Search on MongoDB Community Edition or another self-managed deployment, you install, configure, and operate mongot yourself.
Two products are sold as a service. Pinecone states that it runs on fully managed infrastructure that scales with you. Pinecone states that with BYOC you deploy the Pinecone data plane in your own cloud account (AWS, GCP, or Azure) and get the benefits of a managed service: upgrades, scaling, and maintenance. Pinecone states that Pinecone Local, its in-memory emulator, reached end of life on October 7, 2026. Databricks states that Databricks AI Search, formerly Databricks Vector Search, is built into the Databricks Data + AI Platform and integrated with its governance and productivity tools.
| Deployment pattern | Who runs it | Where the data lives | What to check |
|---|---|---|---|
| Embedded library | Your application process | Local files on the same machine | Whether the same API carries over to a server deployment |
| Single server or container | Your team | Your server | Memory sizing, backups, and the plan for when one machine is not enough |
| Kubernetes, self-managed | Your team, usually with the vendor’s operator or chart | Your cluster | Upgrades, scaling and security, all on your team |
| Vendor-managed in your cloud account (BYOC or hybrid) | The vendor, through an agent or operator in your account | Your cloud account | What the vendor’s control plane can see, and how infrastructure is billed |
| Fully managed service | The vendor | The vendor’s cloud, in a region you choose | Region, data residency, tenant isolation and exit path |
What to Test in a Vector Database
Exact or approximate search comes first. pgvector states that by default it performs exact nearest neighbor search, which provides perfect recall, and that adding an index enables approximate nearest neighbor search, which trades some recall for speed. Approximate indexes are what keep large collections fast, so measure their recall on your own data.
Several products in this guide offer the same graph index. The authors of HNSW present it as a new approach for the approximate K-nearest neighbor search based on navigable small world graphs with controllable hierarchy (Hierarchical NSW, HNSW). pgvector states that an HNSW index has better query performance than IVFFlat (in terms of speed-recall tradeoff), but has slower build times and uses more memory. MongoDB states that MongoDB Vector Search supports approximate nearest neighbor (ANN) search with the Hierarchical Navigable Small Worlds algorithm and exact nearest neighbor (ENN) search. Databricks states that AI Search uses the Hierarchical Navigable Small World (HNSW) algorithm for its approximate nearest neighbor (ANN) searches and the L2 distance metric to measure embedding vector similarity.
Filters change what an approximate index returns. pgvector states that with approximate indexes, filtering is applied after the index is scanned, so if a condition matches 10% of rows, with HNSW and the default hnsw.ef_search of 40, only 4 rows will match on average. Test filtered queries at the selectivity your application uses, not only unfiltered ones.
Hybrid search runs keyword and vector search together. Microsoft defines hybrid search in Azure AI Search as the execution of vector search and keyword search in the same request. Databricks states that AI Search supports hybrid keyword-similarity search, which combines vector-based embedding search with traditional keyword-based search techniques. Test it with queries that contain identifiers, product codes and names, which embeddings alone can miss.
Shared deployments need isolation. The OWASP GenAI Security Project warns that in a multi-tenant environment where different groups share the same vector database, embeddings from one group might be retrieved in response to queries from another group's LLM, potentially leaking sensitive business information. pgvector states that for applications with multiple tenants, sharing an approximate index between tenants means vectors from one tenant can affect recall (and speed) for other tenants. The OWASP GenAI Security Project warns that inadequate or misaligned access controls can lead to unauthorized access to embeddings containing sensitive information, and that the model could then retrieve and disclose personal data, proprietary information, or other sensitive content. OWASP recommends fine-grained access controls and permission-aware vector and embedding stores, with strict logical and access partitioning of datasets in the vector database.
| Capability | What it does | What to test |
|---|---|---|
| Exact nearest neighbor search | Compares the query with every stored vector | Use it as the recall baseline for approximate indexes |
| Approximate index (HNSW, IVFFlat) | Searches a graph or clustered lists instead of every vector | Recall at your target latency, index build time and memory |
| Metadata filtering | Restricts results by fields stored with each vector | Result counts and recall when a filter matches few rows |
| Hybrid search | Runs keyword and vector search together and merges the results | Queries with identifiers, codes and names |
| Multi-tenancy | Separates customers or teams within one deployment | Whether one tenant’s data changes another tenant’s results |
| Access control | Limits who can retrieve which vectors | Whether permissions are enforced at query time |
| Built-in embedding | Creates vectors from text inside the service | Model choice, token cost, and re-embedding when the model changes |
Vendors in Self-Hosted and Managed Vector Databases
Representative vendors include Databricks Vector Search (Databricks AI Search), Elastic / Elasticsearch, Milvus / Zilliz, MongoDB Atlas Vector Search, pgvector / PostgreSQL, Pinecone, Qdrant and Weaviate. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors
The approaches below follow two questions: is vector search the product, or a feature of a database you may already run, and who operates it? Three vendors share the open-source approach.
The managed service approach sells vector search with no servers for you to run. Pinecone describes Pinecone Database as a search engine for AI agents and applications, built for semantic search, knowledge retrieval, and long-term memory at scale.
Three vendors publish an open-source database and sell a cloud service built on it. Weaviate describes itself as an open-source vector database that stores both objects and vectors, allowing for the combination of vector search with structured filtering. Weaviate's GitHub repository lists BSD-3-Clause licenses. Qdrant describes itself as a vector similarity search engine and vector database. Qdrant's GitHub repository lists the Apache-2.0 license. Qdrant states that it is also available as a fully managed Qdrant Cloud, including a free tier. Milvus describes itself as a vector database for storing embeddings and searching them together with structured metadata. Milvus states that it was developed by Zilliz, donated to the LF AI & Data Foundation under the Linux Foundation, and is distributed under the Apache 2.0 license. Zilliz describes Zilliz Cloud as a fully managed vector database and data services built on Milvus. Zilliz states that BYOC lets organizations host applications and data in their own cloud accounts instead of using Zilliz Cloud's infrastructure.
The relational approach keeps vectors in the same tables as the rest of the data. pgvector describes itself as open-source vector similarity search for Postgres that lets you store your vectors with the rest of your data, with ACID compliance, point-in-time recovery and JOINs.
The search engine approach adds vectors to an engine built for keyword search. Elastic states that Elasticsearch functions as a vector database when you store embeddings in dense_vector or sparse_vector fields, and that it can combine vector search with full-text search, structured filters, aggregations, and hybrid retrieval in one engine. Elastic states that ELSER is built-in and requires no external model deployment.
The document database approach indexes vectors inside the operational database. MongoDB states that MongoDB Vector Search lets you search and index your vector data alongside your other MongoDB data.
The data platform approach builds the index from tables already in the lakehouse. Databricks states that you create an AI Search index from a source table, such as a Delta Lake table, and can structure the index to automatically sync when the underlying source table is updated. Databricks states that standard endpoints have a capacity of 320 million vectors at dimension 768, and that storage-optimized endpoints have a larger capacity (over one billion vectors at dimension 768) with a slightly increased query latency of about 250msec.
| Vendor | Approach | Where it fits |
|---|---|---|
| Pinecone | Managed vector database service | Managed service; BYOC on the Enterprise plan; Pinecone Local emulator retired |
| Milvus / Zilliz | Open-source vector database with a vendor cloud | Apache 2.0; Lite, Standalone or Distributed, or Zilliz Cloud and BYOC |
| Qdrant | Open-source vector database with a vendor cloud | Apache-2.0; self-managed, Qdrant Cloud, Hybrid Cloud or Private Cloud |
| Weaviate | Open-source vector database with a vendor cloud | BSD-3-Clause; self-managed on Docker or Kubernetes, or Weaviate Cloud |
| Databricks Vector Search (Databricks AI Search) | Vector search in a data and AI platform | Built into the Databricks platform; indexes sync from Delta Lake tables |
| MongoDB Atlas Vector Search | Vector search in a document database | On Atlas, or self-managed on Community Edition and Enterprise Advanced |
| pgvector / PostgreSQL | Vector search in a relational database | Extension for Postgres 13 and later; exact search by default, HNSW or IVFFlat indexes |
| Elastic / Elasticsearch | Vector search in a search engine | Self-managed, Kubernetes or Elastic Cloud; dense and sparse vectors with keyword search |
Eight vendors across six approaches; Weaviate, Qdrant and Milvus / Zilliz share one. The approaches came first, and each vendor was chosen to represent one. A place in this table cannot be bought. Cloud providers’ own vector services, other database vendors and other open-source projects also compete here, and any of them can make its own case in the Spotlight below.
How much should you budget for Self-Hosted and Managed Vector Databases?
Vector databases are priced on four kinds of meter: a plan minimum or flat fee, usage (reads, writes and storage), dedicated capacity, and the infrastructure you run yourself. Open-source software you run yourself carries no license fee, but its servers, storage and staff time still cost money.
Managed plans start low. Pinecone lists a Builder plan for solo developers and small teams at $20/month flat. Pinecone lists its Enterprise plan at a $500/month minimum usage, with Bring Your Own Cloud (BYOC) included. Weaviate lists a free plan at $0/mo and a pay-as-you-go Flex plan starting at $45/mo. Zilliz lists its Standard plan from $0/month (Serverless) and from $126/month (Dedicated).
Memory and capacity drive the larger bills. MongoDB states that MongoDB Vector Search holds the entire index in memory. Size servers for the whole index, its replicas and headroom for growth.
Under a bring-your-own-cloud arrangement, ask for two figures: the vendor’s fee, and the infrastructure you pay your cloud provider for.
Embedding is a separate cost. Every record is embedded once on load and again whenever you change the embedding model, so price a full re-embed before you commit to a model.
| Meter | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Plan minimum or flat fee | A monthly commitment, usually credited against usage | Plan tier | Paying for a tier above what production needs |
| Usage | Reads, writes and stored data | Traffic and collection size | A full re-embed multiplies writes |
| Dedicated capacity | Reserved nodes or compute units | Vector count, dimension and replicas | Sizing for peak and paying for it all month |
| Self-hosted infrastructure | Servers, memory, disks and staff time | Index size in memory and replicas | Underestimating upgrade and on-call work |
| Embedding | Tokens sent to an embedding model | Corpus size and model changes | A model change forces a full re-embed |
How long does implementation take for Self-Hosted and Managed Vector Databases?
A vector database rollout is mostly a data project: chunking, embedding, loading, and keeping the index in step with the source system. Choose the embedding model before the database, because changing it later means re-embedding every record.
Estimate vector count, dimension and growth. Write the filters the application needs, and collect real queries with known good answers.
Load a representative sample into two finalists, including the database you already run if it supports vectors. Record recall, latency, memory and projected cost.
Build the sync from source systems, set tenant and permission boundaries, and decide who operates the database and how it is backed up.
Monitor recall on the test set, index size and cost. Re-run the test set before any change of embedding model or index settings.
What should you ask vendors about Self-Hosted and Managed Vector Databases?
Ask each vendor to answer in writing and to show the answer on your data, at your vector count, with your filters.
-
Does the data you want to search already live in a database or search engine that supports vector search?Yes Test that product’s vector search first, at your expected vector count, before adding a separate database.No Shortlist dedicated vector databases and compare them on filtering, hybrid search and operations.
-
Must the vectors stay on infrastructure you control?Yes Shortlist products with a self-managed or bring-your-own-cloud option, and price the staff time to run them.No Compare managed services on region, isolation, plan minimums and exit path.
-
Will many customers or teams share one deployment?Yes Test tenant isolation and query-time permission filtering before performance.No Weight recall, latency and cost at your production vector count.