Pinecone adds Gemini Embedding 2 to managed document search

By Sophie Lindqvist

Pinecone's Documents API now supports Google's Gemini Embedding 2 at $0.20 per 1 million tokens, Pinecone announced on November 12, 2026. The launch gives teams a managed route from plain text to searchable embeddings through Vertex AI. Applications no longer need a separate Google Cloud account, API key or embedding pipeline.

Pinecone said the service invokes Gemini Embedding 2 on every document write and search query, then retries failed calls. Applications send plain text, and Pinecone handles the model work. Embedding usage appears on Pinecone's invoice rather than a separate Google bill.

This moves a model-provider step into the database path. For teams building retrieval systems, that removes one pipeline to design, monitor and bill, while leaving text and queries under the application's control.

What changes when embedding runs inside the database?

Pinecone's documentation describes embeddings as vectors that represent meaning. With the managed model, one text field can support similarity-based retrieval and keyword search. The application does not have to generate a vector before sending the content to Pinecone.

Pinecone also writes each document's text and embedding together. The company says this gives semantic and full-text search the same version of that document. In practice, the two search modes read from one managed document state instead of relying on separately generated vectors.

For search teams, the practical shift runs from a separately managed provider integration to an option inside the database. Pinecone invokes Google and retries failed calls, while customers decide what text and queries to submit.

Takeaway: The integration removes an embedding-building step, but it does not remove the need to design search carefully.

How are index settings fixed?

At index creation, customers select gemini-embedding-2 in a text field's embed configuration. Pinecone's model documentation lists 3,072 dimensions by default. Any integer from 128 through 3,072 is supported, and cosine is the only available similarity metric.

The model accepts up to 8,192 tokens per input. Pinecone fixes the model, dimension and metric when the index is created. The company describes these choices as infrastructure decisions, not parameters that teams can adjust after deployment.

That makes index creation the point for deciding how Pinecone will represent and compare text. Teams cannot tune those three settings on a live index through the documented configuration, so they need to settle the architecture first.

Takeaway: Model, dimensions and metric belong in the initial architecture review, because the index locks them in.

How do task types shape embeddings?

Pinecone's task types control how Gemini Embedding 2 embeds text for different workloads. General retrieval uses retrieval_document for stored text and retrieval_query for searches by default. Classification, clustering and sentence-similarity tasks treat stored text and queries alike for text comparisons.

Other query settings cover question answering, fact verification and code retrieval. A single search can override the field's default query task. That lets one search request select a different task without changing the field's stored default.

Pinecone inserts Gemini's required task instruction automatically. Users submit plain text instead of constructing model-specific prompts. The integrated-embedding documentation also fixes a field's write task when the index is created.

Those instructions add about six or seven tokens to every value or query. Pinecone says they count toward the model's 8,192-token input limit. The overhead therefore applies to both indexed content and incoming searches.

Takeaway: Task selection is mostly managed, but teams still choose the default and can adjust individual queries.

What does the service cost, and who can use it?

Pinecone charges for embedding tokens on both writes and queries. That ties consumption to indexing and search activity rather than to one operation alone. The published price is $0.20 per 1 million tokens.

The Builder plan normally includes 10 million Gemini Embedding 2 tokens per month. Pinecone says that allowance doubles to 20 million through December 31, 2026. The model requires a Builder, Standard or Enterprise plan, and it is unavailable on Starter.

One embedding request can contain up to 96 inputs. A search query is limited to 10,000 bytes. These request and payload limits sit alongside the model's 8,192-token input ceiling, so teams must shape both indexing and query traffic around those boundaries.

Takeaway: Usage follows both stored content and live search, with a temporary Builder allowance that doubles the normal monthly limit.

What works at launch?

The first release supports only serverless Documents API indexes. Pinecone excludes BYOC, pod-based and Records API indexes from this release. Customers can use the console or Python SDK 10.3.0, while Node.js and Go SDK support are marked “coming soon” in Pinecone's announcement.

Pinecone can convert existing vector indexes to Documents API indexes without re-embedding. The catch is that customers cannot yet add integrated embedding to a converted index. Pinecone's current workflow therefore uses new Documents API indexes for the feature.

That restriction separates the managed launch from a direct upgrade path for existing deployments. Teams starting a new text-search project can use the release as configured, while established vector-index customers cannot yet add the integrated model after conversion.

Takeaway: Existing customers can convert an index, but they cannot yet use Gemini Embedding 2 inside that converted index.

What happens to multimodal search?

Google's March 10, 2026 model announcement called Gemini Embedding 2 “natively multimodal.” Google said the model maps text, images, video, audio and documents into one embedding space. Pinecone's launch, however, covers only text and lists multimodal support for a later release.

Pinecone says its text service supports more than 100 languages. That gives the initial integration broad language coverage, but it does not yet expose the model's ability to place images, video, audio and documents in the same space.

Processing is global. Pinecone says Google can handle each request in any of its regions, but the company does not guarantee a particular region. The policy offers geographic flexibility without firm regional placement for an individual request.

Takeaway: Pinecone has a route toward multimodal retrieval, while the launch remains a text-only, globally processed service.

What does this change for search teams?

For teams operating search infrastructure, Pinecone's release moves embedding from a separately managed provider integration into the database path. Its immediate fit is new text-search deployments on paid plans. Global-only processing and the converted-index restriction define the scope of the initial service.

The practical effect is a shorter setup path. Pinecone now handles Google model access, vector generation, document storage, and semantic and full-text retrieval within one service. That can remove a separate pipeline, but customers still must choose dimensions, metrics, task types and write behavior before index creation.

Pinecone's next concrete milestones include multimodal support and Node.js and Go SDKs. Integrated embedding for converted indexes remains unavailable, so the service fits new, paid, serverless text indexes first, not a wholesale replacement for every existing search setup.

Takeaway: Pinecone's next proof points are multimodal support, Node.js and Go SDKs, and a working path for adding Gemini Embedding 2 to converted indexes.