Elasticsearch Vector Database project overview

The Elasticsearch Vector Database Serverless project type is optimized for vector workloads. Compared with the general-purpose Elasticsearch project type, it uses a vector-tuned default configuration, a hardware profile suited to embeddings, streamlined access to inference, and a pricing model built for vector storage and search.

Use the Vector Database project type when embeddings and similarity search are central to your application, for example RAG, recommendations, semantic search, hybrid search, or multimodal search.

A Vector Database project gives you the same Elasticsearch vector search capabilities as other project types, plus Elastic Cloud Serverless defaults and project settings aimed at embedding storage, inference, and similarity or hybrid search.

Like any Elasticsearch index, you can store embeddings alongside standard field types in the same documents, such as keyword, text, numeric types, date, boolean, geo fields, nested, and more. Combine similarity search with filters on that metadata using Query DSL queries such as bool, range, and terms. Combining vectors and structured fields in one engine is an advantage over a typical dedicated vector store.

Indices use the index.mode: vectordb_document vector index mode automatically. It applies storage, indexing, and merge defaults tuned for similarity search on dense vectors, so you get efficient embedding storage and approximate kNN search without configuring each setting yourself. This is the only index mode supported in this project.

Index-level vector tuning is managed for you. You can balance search capacity, latency, and cost by adjusting Search Power settings.

The project hardware profile is also tuned for vector workloads.

Vectors are stored compressed to reduce storage size. If _source is enabled for vector fields, it keeps the values you indexed. Retrieving those same vectors with the fields parameter can return slightly different values because of compression. For example, 3.0 might read back as 2.9953578. That difference is expected; it does not mean your data is corrupted.

The Vector Database project type favors workloads where you ingest and embed data, then serve similarity or hybrid queries repeatedly. Aggressive segment merging improves query speed for data that changes infrequently, which is a common pattern for knowledge bases, product catalogs with semantic search, and RAG document stores.

Vector Database projects are set up for embedding workloads: use managed workflows with semantic_text or semantic fields to generate embeddings in Elasticsearch, or bring your own vectors and search them with a kNN query.

Similar to other Elastic Cloud Serverless projects, Elastic manages the infrastructure, scaling, and upgrades. You create a project, get an endpoint, and start indexing and querying without sizing nodes for vector RAM yourself. Each project can store up to 1 TB of data.

Tip

You can check your dataset size with the Get index information API.

Billing is based on storage, search, ingest, and infrastructure, rather than the compute-based VCU model used by Elasticsearch Serverless projects. That means costs follow how much you store and the search capacity you reserve with Search Power, not fluctuating compute or query volume, which is a better fit for embedding data you query often. Refer to Elasticsearch Vector Database billing dimensions for details.

Both the Elasticsearch Vector Database and the Elasticsearch project types support vector search. Vector Database gives you the full power of the core Elasticsearch capabilities (the same query APIs, mappings, filters, and hybrid or semantic retrieval) with targeted defaults, hardware, and pricing tuned for embedding-driven workloads.

Choose Vector Database when embeddings and similarity search are the primary workload, and you don't need extra features like time series data support, search application management, or custom ML nodes.

Choose the Elasticsearch project type when you need general-purpose data storage and search, including mixed lexical, time series, and analytics workloads, Kibana search tooling such as Query Rules UI, or the ability to run custom models on ML nodes. You might also prefer the Elasticsearch project type if you are an existing Elasticsearch or OpenSearch user.

Use case Fit Why
RAG and question answering Strong Retrieve passages from documents, wikis, tickets, or knowledge bases and pass them to an LLM. Hybrid search combines semantic similarity with keyword matching when queries mix natural language with exact terms, IDs, or product names.
Discovery and recommendations Strong Find related products, articles, or other items by similarity when keywords alone are not enough. Use hybrid ranking when you also need lexical or attribute matches in the same result set.
Multimodal search Strong Search across images, audio, video, or text with embeddings from a multimodal model.
Duplicate detection, fraud, and anomaly detection Strong Compare embeddings to find near-duplicates, suspicious matches, or unusual patterns at scale.
Long-term memory for LLMs Strong Store facts, chat turns, or summaries so an assistant can retrieve relevant past context by meaning, optionally combined with keyword filters on metadata.
Full-text or keyword search without vectors Prefer the Elasticsearch project General-purpose defaults suit lexical search, filters, and document-centric analytics.
Log, event, or other time series search Prefer the Elasticsearch project General-purpose defaults suit write-heavy, frequently updated time series data.
Note

Vector Database projects only support the vector index mode. Time series index mode and LogsDB index mode are not supported. Use the Elasticsearch project type for those workloads. Data streams are supported when their backing indices use vector index mode.

Ready to try the Vector Database project type? For a hands-on, step-by-step walkthrough with a client, sample data, and semantic, hybrid, and ES|QL searches, follow Elasticsearch vector and full-text search in 10 minutes.

For an overview of creating a project and the ingest and search options available, refer to Get started with the Elasticsearch Vector Database project type.

Edit page