Setting the file. One moment.
Chapter 45 · Amazon Opensearch Service
Subchapter 45.36
references/search-sparse-vector-models.mdMarkdown5 KBView on GitHub
This document lists the available models for Sparse Vector (Neural Sparse) Search in OpenSearch, categorized by deployment mode.
Deploying sparse models directly on OpenSearch Nodes.
Note: Running sparse encoding models on CPU OpenSearch Nodes is generally not recommended for high-throughput production due to latency. CPU OpenSearch Nodes are best suited for tokenizers in Doc-only mode search, or low-traffic/dev sparse encoding inference.
| Model Name | Type | Description | Recommended Use |
|---|---|---|---|
amazon/neural-sparse/opensearch-neural-sparse-tokenizer-v1 | Tokenizer | Neural sparse tokenizer with IDF-based token weights (defaults to 1 if IDF not provided). | Search Phase (Doc-only mode) |
amazon/neural-sparse/opensearch-neural-sparse-tokenizer-multilingual-v1 | Tokenizer | Multilingual neural sparse tokenizer with IDF-based token weights (defaults to 1 if IDF not provided). | Search Phase (Multilingual Doc-only mode) |
| Model Name | Type | Description | Recommended Use |
|---|---|---|---|
amazon/neural-sparse/opensearch-neural-sparse-encoding-v1 | Sparse Encoder | Neural sparse encoding model (bi-encoder style). | Dev / Low traffic |
amazon/neural-sparse/opensearch-neural-sparse-encoding-v2-distill | Sparse Encoder | Distilled v2 sparse encoding model. | Dev / Low traffic (or GPU for prod bi-encoder) |
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v1 | Doc Encoder | Document-side sparse encoder for doc-only setups. | Dev / Low traffic |
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v2-distill | Doc Encoder | Distilled doc encoder v2. | Dev / Low traffic |
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v2-mini | Doc Encoder | Smaller “mini” doc encoder v2. | Dev / Low traffic / cost-sensitive experiments |
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v3-distill | Doc Encoder | v3 distilled doc encoder. | Dev / Low traffic (or GPU for prod doc-only) |
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v3-gte | Doc Encoder | v3 GTE-based doc encoder. | Dev / Low traffic (or GPU for prod doc-only) |
amazon/neural-sparse/opensearch-neural-sparse-encoding-multilingual-v1 | Sparse Encoder | Multilingual neural sparse encoding model. | Dev / Low traffic (or GPU for prod multilingual) |
Not Supported. Custom or fine-tuned sparse encoding models cannot be deployed on OpenSearch Nodes. You must use a SageMaker GPU Endpoint.
Deploying sparse models on AWS SageMaker with GPU acceleration is the recommended strategy for:
The models listed in 1.1. For tokenizers, it’s recommended to get deployed on OpenSearch nodes. For deep learning models, the recommended instance type is ml.g4dn.xlarge or ml.g5.xlarge.
If you have trained a custom or fine-tuned sparse encoding model, you must deploy it using a SageMaker GPU Endpoint. This deployment mode supports custom model logic and weights that are not available in the pre-trained registry.
In this mode, you decouple ingestion and search compute.
amazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v3-gteamazon/neural-sparse/opensearch-neural-sparse-encoding-doc-v3-distillamazon/neural-sparse/opensearch-neural-sparse-tokenizer-v1In this mode, query processing is heavy and requires inference.
amazon/neural-sparse/opensearch-neural-sparse-encoding-v2-distillamazon/neural-sparse/opensearch-neural-sparse-encoding-v2-distillamazon/neural-sparse/opensearch-neural-sparse-encoding-multilingual-v1amazon/neural-sparse/opensearch-neural-sparse-tokenizer-multilingual-v1