pinecone-sparse-english-v0

Overview

pinecone-sparse-english-v0 is a learned sparse embedding model built on the innovations of the DeepImpact architecture, estimating lexical importance of tokens by leveraging their context.

Performance

• Outperforms BM25 by up to 44% (average 23%) on TREC Deep Learning Tracks
• Measured by normalized discounted cumulative gain (NDCG)@10
• Superior keyword matching with contextual understanding
• Optimized for English language retrieval

Technical Approach

• Based on DeepImpact architecture
• Context-aware token importance estimation
• Learned sparse representations
• Unlike BM25, leverages semantic context not just term frequency

Hybrid Search Performance

• Combined with dense retrieval and reranking (cascading retrieval)
• Achieves up to 48% better performance than sparse or dense alone
• Optimal for precision-critical applications

Integration

• Native Pinecone Inference API support
• Sparse index compatibility
• Works with Pinecone Text Client
• Seamless hybrid search integration

Use Cases

• High-precision keyword search
• Hybrid semantic + lexical search
• Domain-specific retrieval requiring exact matches
• Cascading retrieval pipelines
• Enterprise search applications

Pricing

Included with Pinecone API usage, priced per inference call