Introducing support for sparse-dense embeddings for better search result

Pinecone has introduced public preview support for sparse-dense embeddings in search indexes, aiming to enhance search experience by integrating both semantic and keyword search capabilities through a hybrid approach. This advancement allows users to achieve more relevant search results by utilizing the strengths of both dense vector embeddings, which capture semantic meanings, and sparse vector embeddings, which highlight term-level importance. The new index supports various large language models and accommodates a range of use cases, including multimodal search and boosting, without the need for separate configurations. Leveraging models like SPLADE for sparse vector generation, the upgraded index provides flexibility in combining sparse and dense elements, thus surpassing traditional methods like BM25 in relevance and ranking. With a user-friendly REST API and Python SDK, users can quickly implement and experiment with this technology, which is designed to transform the way high-value information is accessed by balancing term occurrence with semantic nuances.