Pinecone Adds Integrated Embedding and Search to the Documents API
- Cargo
- Machine Learning
- Shipped
- Shipper
- Richard Williams
- Readers
- 934
Pinecone combines text embedding and search in one request, with semantic, sparse and BM25 scoring plus support for several embedding models.

Pinecone’s Documents API now handles text embedding and search through the same interface. For teams building search applications, the practical change is fewer components to operate: documents and queries can be embedded within Pinecone, while semantic and full-text search use the same text field. That removes the need to run a separate embedding service and coordinate its keys, retries and monitoring with the search pipeline.

The feature is available on serverless indexes. At launch, Pinecone listed three models: llama-text-embed-v2, multilingual-e5-large and pinecone-sparse-english-v0. An update adds support for OpenAI’s text-embedding-3-small and text-embedding-3-large, along with Google’s gemini-embedding-2. Teams can therefore use several hosted embedding options through the Documents API rather than maintaining a separate call to an embedding provider.
One text field, several ways to search
Integrated embedding is configured in an index schema. A text field can have an embedding model and full-text search settings; Pinecone applies the selected embedding model to documents and queries that use that field. The index description records each field’s model, vector dimension and distance metric. For English sparse retrieval, a sparse_embed setting adds Pinecone’s sparse model to the same field.

Search requests can rank results in three ways. embed uses vector similarity to find passages with related meaning. sparse_embed uses learned keyword weights, while text uses BM25 keyword matching. Applications choose a scoring method per request, and can combine semantic ranking with a full-text filter. For example, a query about fixing an error can be restricted to documents containing the exact phrase ERR-4031, then ranked by semantic relevance.
This arrangement also keeps document and query vectors aligned: Pinecone embeds both with the model configured for the field. When text changes, the write request triggers a new embedding and replaces the previous vectors. Updates to other fields don’t consume embedding tokens. If embedding fails, Pinecone writes nothing from that request, avoiding a partially updated document.

Model limits and configuration
llama-text-embed-v2 is the default semantic model. It accepts up to 2,048 tokens and supports 26 languages, including Spanish, Chinese and Japanese. Its available dimensions are 384, 512, 768, 1,024 and 2,048, with 1,024 as the default.
Microsoft Research’s open-source multilingual-e5-large supports shorter multilingual inputs, up to 507 tokens, at 1,024 dimensions. pinecone-sparse-english-v0 is limited to English; it accepts 512 tokens by default and up to 2,048, producing sparse vectors without a fixed dimension. Pricing for the models is listed on Pinecone’s pricing page.

These limits shape model choice: longer passages fit the default model’s input window, while E5 has a shorter maximum input and a fixed dimension. Sparse retrieval provides a separate option for keyword-focused matching. The announcement doesn’t include comparative quality scores, so teams evaluating relevance will still need to assess models against their own documents and queries.
Text storage and migration
Teams that want semantic vectors without retaining the original text can set "store_text": False on a field. Pinecone embeds incoming text and discards it, keeping the vectors instead. That field can’t support full-text search or $match_* filters, and requesting it through include_fields returns an error. Search results can still include the document ID, score and other stored fields.

Existing data can be moved by creating a Documents API index with an embedding model on its text field, then upserting the text. Pinecone embeds each document during the backfill, using the same configured model that will process later queries. The backfill is billed per token at the listed model rates.
For teams currently using the Records API, the Documents API retains a text-in, text-out workflow and adds full-text search. Pinecone says it is working on support for moving Records API indexes to the Documents API; that migration path is not yet described as available.
Latency and availability
Combining embedding and search also removes a separate network call. In a Pinecone test from a client in US East, using llama-text-embed-v2 at 20 queries per second, the integrated request had a median latency of 26 ms. The comparison workflow, which first called the standalone embedding API and then searched with the resulting vector, had a median of 92 ms. Those figures describe one test setup, not a general latency guarantee.
The feature works through the Pinecone API version 2026-07, Python SDK 10.1.0, TypeScript SDK 9.1.0 and Go SDK 7.1.0 or later. Pinecone has also outlined further work: server-side fusion to combine full-text and semantic results into a ranked list, integrated reranking, embedding during bulk import, and processing within a customer-selected geography. The immediate change is simpler deployment; fusion and the other planned additions could reduce more application-side work once available.