opensearch-neural-sparse-encoding-doc-v2-distill
| Property | Value |
|---|---|
| Parameter Count | 67M |
| License | Apache 2.0 |
| Paper | Research Paper |
| Average NDCG@10 | 0.504 |
| FLOPS | 1.8 |
What is opensearch-neural-sparse-encoding-doc-v2-distill?
This is a learned sparse retrieval model specifically designed for efficient document search in OpenSearch environments. It represents a significant advancement in neural search technology, encoding documents into 30,522-dimensional sparse vectors while maintaining high search relevance with reduced computational overhead.
Implementation Details
The model utilizes a distilled architecture based on DistilBERT, achieving efficient document encoding through sparse vector representation. It's trained on diverse datasets including MS MARCO, WikiAnswers, and various StackExchange collections, making it robust for real-world applications.
- Inference-free retrieval capability for enhanced efficiency
- Optimized performance with only 67M parameters
- Competitive NDCG@10 score of 0.504 across benchmark datasets
- Significantly reduced FLOPS (1.8) compared to predecessor models
Core Capabilities
- Efficient document encoding to sparse vectors
- Weight-based token importance calculation
- Seamless integration with OpenSearch's Lucene inverted index
- Optimal balance between search relevance and computational efficiency
Frequently Asked Questions
Q: What makes this model unique?
The model's key distinction lies in its ability to perform inference-free retrieval while maintaining competitive search relevance. It achieves this through a distilled architecture that's significantly smaller than its predecessors while delivering comparable or better performance.
Q: What are the recommended use cases?
This model is ideal for large-scale document retrieval systems where efficiency and accuracy are crucial. It's particularly well-suited for enterprise search applications, content recommendation systems, and any scenario requiring fast, accurate document retrieval with reasonable computational resources.
