Elastic Search 8.15: Accessible semantic search with semantic text and reranking

Elasticsearch 8.15 introduces semantic text and semantic reranking — two powerful ways to bring natural language search into your search experiences. 

We continue to make vector search more accessible with semantic text — a new field type and corresponding semantic query type that unlocks a streamlined approach to vector search for text. In particular, semantic text handles automated chunking, so users with long text fields can rest easy knowing they have full multi-embedding coverage and handling done under the hood. Vector search customers can just set a semantic text field to get started with model coordination and embedding generation handled automatically by the inference API.

For users who want semantic search without vectors or to further fortify their vector or hybrid search, 8.15 also introduces semantic reranking. This feature can dramatically improve search results, especially natural language search, by applying powerful semantic similarity models to refine the Top N results of a search query. This addition is a potentially powerful improvement for any search system, but it's especially valuable for users seeking to improve natural language search capabilities without incorporating vector embeddings.

Elasticsearch's open inference API continues to give our users even more model flexibility with additional third-party providers. 8.15 introduces:

• Google AI Studio (embeddings and completions) and Vertex AI (embeddings and reranking)

• Mistral (embeddings)

• Amazon Bedrock (embeddings and completions)

• Anthropic (completions)

All of these model providers are easily configured as an inference endpoint by providing a simple API key and then can be used with retrievers for simple and easy composability of the search query. 

Elasticsearch's native Learning to Rank and query rules features are generally available as of 8.15, offering users even finer control over their search relevance and results.

• Learning to Rank is built directly into Elasticsearch and expands the search relevance tuning pipeline of our search users, so they can rerank results with an LTR model trained on the features that matter to them. This runs directly in Elasticsearch and enables more fine-tuned and context-aware search results that can adapt to complex search and user behavior patterns. And as of 8.15, you can also rescore collapsed results, which means you can use LTR to rerank your field collapsed results.

• Query rules introduce the ability to apply business rules to the search set even with the use of machine learning models for hybrid search. Users can specify the desired documents to pin to the top of search results when the defined conditions are met — enabling search results that are tailored by query, user segment, and more, and it works with any combination of hybrid search and RRF.

• Playground now supports Google Gemini!

Read about these capabilities and more in the release notes.

Existing Elastic Cloud customers can access many of these features directly from the Elastic Cloud console. Not taking advantage of Elastic on cloud? Start a free trial.