Elasticsearch Vector Search

Elasticsearch Vector Search extends the well-known Elasticsearch engine with first-class dense and sparse vector storage, ANN search (HNSW), and hybrid retrieval that blends BM25 keyword scoring with semantic similarity in a single query. It is aimed at engineering teams building retrieval-augmented generation (RAG) pipelines, semantic search over enterprise documents, product search, and recommendation systems who want a vector database that also handles full-text, filters, faceting, geo, and time-series in one cluster instead of stitching together specialty stores. The platform ships a `semantic_text` field type that automatically chunks documents and calls a configured embedding provider (OpenAI, Cohere, Hugging Face, Mistral, Azure AI, Bedrock, Vertex AI, or Elastic's built-in ELSER sparse model) at index time, so teams do not have to hand-roll ingestion pipelines. Better Binary Quantization (BBQ) cuts vector memory by up to ~95% for large corpora, and native filters run alongside vector search without collapsing recall. Elastic also provides an AI Playground for iterating on retrieval strategies, LangChain and LlamaIndex integrations, and a reranking API (including Learn-to-Rank) that plugs into RAG stacks. Typical workflows: ingest via connectors or the Bulk API, generate embeddings inline with `semantic_text`, run a hybrid kNN + BM25 query with metadata filters, optionally rerank, then hand the top-k passages to your LLM. Deployment ranges from Elastic Cloud Serverless (fully managed, autoscaling) through Cloud Hosted on AWS/Azure/GCP to fully self-managed Kubernetes or air-gapped installs.