Pinecone now hosts Google's Gemini Embedding 2 inside its own index

Running retrieval on a hosted embedding model has always come with two accounts, two keys and two invoices. One for the vector database, one for the embedding provider. Pinecone shortened that this week. Product managers Gavin Johnson and Manish Talreja announced the change. Google's Gemini Embedding 2 is now part of Pinecone's integrated embedding.

What changed

Integrated embedding is Pinecone's feature that runs the embedding call inside its own index. A team sets a model on a text field when creating a Documents API index. The database then embeds every write and every query for that field with the chosen model. Until now, the choices were Pinecone-hosted models that launched with the feature. Gemini Embedding 2 is the first large third-party model Pinecone has added.

Pinecone serves Gemini Embedding 2 from Google Vertex AI. It retries failed calls so a team does not have to build its own retry logic. There is no Google Cloud account or Google API key to manage. The tokens appear on the Pinecone invoice at Google's list price.

How it is used

A team creates an index with a schema and sets Gemini Embedding 2 on the text field. The post shows a Python example. It embeds a body field at 1,536 dimensions and turns on full-text search on the same field. The same field then serves both semantic and keyword search. Both views see the same version of each document.

A single call writes the document and its embedding together. A query sends plain text and the database embeds and searches in one trip. The code change for a team already on a Pinecone Documents API index is small.

Pricing and limits

Gemini Embedding 2 needs a Builder plan or above on Pinecone. The post states one limit explicitly. Until 31 December 2026, the Builder plan gets twice its included tokens for Gemini Embedding 2: 20 million tokens per month instead of 10 million. The price per token is Google's own list rate, billed by Pinecone.

Pinecone does not quote latency or recall figures for Gemini Embedding 2 in the post. The company describes the model as Google's newest embedding model, and points to Google's own documentation for specifics.

Why it matters for an AI engineering team

Retrieval stacks accumulate vendors. A team on Pinecone usually also has a key with a separate embedding provider. There is a second invoice, and a thin wrapper to handle the API calls and the retries. Each of those is a small cost on its own. Together they are the work a team does before it starts on relevance tuning or prompt design.

Moving the embedding call inside the index removes one vendor relationship for teams that pick Gemini Embedding 2. It also makes write and query embeddings consistent, because the same code path runs them. The Builder-plan credit through December reads as a promotion to pull trials onto the combined path.

The change is narrow. A team already happy with its own embedding pipeline gets no new capability here. A team that picked Pinecone because it keeps retrieval simple gets one less thing to manage. That is the shape of most retrieval work today: not a new model, but one fewer account, one fewer key, one fewer bill.

Source

Pinecone integrated embedding now supports Google's Gemini Embedding 2, by Gavin Johnson and Manish Talreja at Pinecone.