Elastic launches Elasticsearch Vector Database for AI search

Elastic N.V. just rolled out a purpose-built vector database, and it’s aimed squarely at the AI infrastructure layer that every enterprise is scrambling to build. The Elasticsearch Vector Database, announced on September 11, is a serverless product running on Elastic Cloud Serverless that’s designed to handle vector search at massive scale, think hundreds of billions of vectors.

What the product actually does

Vector databases are the memory layer behind modern AI applications. When a large language model needs to retrieve relevant information before generating a response, a process called retrieval-augmented generation (RAG), it queries a vector database to find semantically similar content. Elastic’s new offering is built specifically for this use case, along with powering AI agents that need to search across large knowledge bases.

The product ships with pre-optimized configurations for indexing, embedding, chunking, and hybrid retrieval. In plainer terms: developers don’t have to spend weeks tuning the system before it works well with their AI pipelines.

One of the more notable technical features is support for multiple data types, including text, image, and multi-modal vectors, all consolidated into a single index. That means a developer building a product search engine could query across product descriptions and product images simultaneously without maintaining separate databases.

Elastic also integrated managed GPU inference through its Elastic Inference Service, which supports both native models and third-party options like those from Jina AI. The practical implication: teams can run embedding and reranking models without spinning up separate model servers.

Hybrid search is another selling point. The database enables both vector similarity queries and traditional keyword searches on the same index.

The compression trick

Perhaps the most technically impressive claim involves Elastic’s proprietary Better Binary Quantization (BBQ) and DiskBBQ technologies. These achieve up to 32x memory reduction while maintaining strong recall and speed.

The VectorDB index mode and automatic calibration features underlying this launch became generally available in August 2026, so the foundation has been in production for at least a month before the full product announcement.

On pricing, Elastic opted for a model based on data volume and search capacity rather than traditional compute unit pricing. Compute-based pricing models often produce unpredictable bills because usage patterns fluctuate. Volume-based pricing gives finance teams something they can actually forecast.

A crowded field gets another entrant

Elastic isn’t entering an empty market. Purpose-built vector database startups like Pinecone, Weaviate, Qdrant, and Milvus have been attracting venture capital and enterprise customers for several years now. Meanwhile, incumbent database companies like MongoDB and PostgreSQL (via pgvector) have added vector search capabilities to their existing products.

Elastic’s argument is essentially that enterprises already using Elasticsearch for search and observability shouldn’t need a separate specialized database for AI workloads. Consolidation is the pitch: one platform for keyword search, vector search, and hybrid queries.

ESTC stock was trading around $85 in mid-September 2026, with analysts reportedly upgrading the company amid its AI infrastructure push.