A string field in a document schema can now be configured for both functions, so a client sends text rather than precomputed vectors on writes and searches.
In Pinecone's example, a field uses full-text search options and the llama-text-embed-v2 embedding model. On a write, Pinecone embeds the text, stores the resulting vector next to it, and builds the text index in the same request. If embedding fails, the company says the request is not written. On a search, the API embeds the text query using the field's configured model.
The full-text implementation stores raw text in Pinecone's slab architecture and uses a Tantivy inverted index with BM25 scoring; it also supports text-match filters such as $match_phrase. Pinecone says its own tests found a one-call text search about 3.5 times faster at the median than calling its standalone embedding API followed by vector search, chiefly by removing a network round trip.
The post does not state pricing, plan or regional availability, throughput limits, or independent latency and retrieval-quality results. It describes Pinecone's implementation and internal testing rather than a comparative benchmark.