opensearch-muvera-plugin

An Apache-2.0 Java plugin for self-managed OpenSearch, published in the aws-samples/AI-search-with-amazon-opensearch-service repo, that computes MUVERA fixed dimensional encodings server-side.

• Repo: https://github.com/aws-samples/AI-search-with-amazon-opensearch-service/tree/muvera-late-interaction/muvera-late-interaction

---

What It Adds

• muvera ingest processor — encodes a document’s multi-vectors into one FDE stored in a knn_vector field.
• muvera_query search processor — encodes the query the same way.

The FDE drives first-stage ANN; candidates are then reranked with exact MaxSim via k-NN’s lateInteractionScore. The prebuilt zip targets OpenSearch 3.8.0 only (OpenSearch requires an exact version match); other versions need a source build with JDK 21. On managed Amazon OpenSearch Service, where custom plugins can’t be installed, the same encodings come from FastEmbed on the client instead.

Benchmarks with ColBERTv2 and ColModernVBERT are in Scaling Late Interaction Retrieval in OpenSearch with MUVERA.

Related Concepts

• MUVERA · Late Interaction · ColBERT · ColPali · Reranking

Related Topics

• Late Interaction in OpenSearch

Related Articles

• Scaling Late Interaction Retrieval in OpenSearch with MUVERA