Pinecone Adds Gemini Embedding 2 to Its Documents API

Pinecone’s Documents API can now generate embeddings with Google’s Gemini Embedding 2, letting teams send text directly for indexing and search. Pinecone handles the model calls for both document writes and queries, so developers don’t need to run a separate embedding service or manage Google credentials. The tradeoff is that the integration currently covers text fields on serverless Documents API indexes, while several other index types and multimodal inputs remain unavailable.

Gemini Embedding 2 joins the Pinecone-hosted models that were already available through integrated embedding. Teams configure the model on a text field, then send plain text to Pinecone. The service embeds each document when it is written and embeds search queries when they arrive. Embedding tokens appear on the Pinecone invoice at Google’s list price, rather than on a separate Google bill.

The Builder plan includes 10 million Gemini Embedding 2 tokens per month. Through December 31, 2026, that allowance is doubled to 20 million tokens per month. The model requires Builder or a higher plan, so the token promotion does not make the feature available on lower tiers.

One service handles embedding and search

With a direct Google embedding integration, an application generally needs to call the model when it writes a document and again when it prepares a query. That means handling credentials, rate limits, retries and a separate provider bill. Pinecone’s managed path puts those calls inside its indexing and search operations: the client sends text, and Pinecone embeds it before storing or searching.

Document text and its embedding are written together. This keeps the searchable text and vector tied to the same document version, which matters when content changes and both keyword and semantic search are in use. Pinecone serves Gemini Embedding 2 through Google Vertex AI and retries failed model calls, reducing the integration work teams need to maintain themselves.

The API can also combine semantic search and full-text search over the same field. In the example below, the body field is configured for both, with Gemini Embedding 2 producing 1,536-dimensional vectors:

from pinecone import Pinecone

pc = Pinecone(api_key="YOUR_API_KEY")

pc.indexes.create(
    name="support-articles",
    schema={
        "fields": {
            "body": {
                "type": "string",
                "full_text_search": {},
                "embed": {
                    "model": "gemini-embedding-2",
                    "dimension": 1536,
                },
            }
        }
    },
)

index = pc.index("support-articles")

index.documents.upsert(
    namespace="articles",
    documents=[
        {"_id": "kb-205", "body": "Refunds take five business days."},
    ],
)

A query can be sent as text, with a task type that describes the search intent:

results = index.documents.search(
    namespace="articles",
    top_k=5,
    score_by=[
        {
            "type": "embed",
            "fields": ["body"],
            "query": "how long do refunds take",
            "parameters": {"task_type": "question_answering"}
        }
    ],
    include_fields=["body"],
)

Dimensions and task settings

Setting dimension to 1,536 is optional. If it’s omitted, Pinecone uses the model’s full vector size. The integration accepts any dimension supported by the model, which can help teams retain the vector size they already use when configuring new indexes.

Gemini Embedding 2 expects a short instruction that identifies whether an input is a document or a query. Pinecone adds the appropriate instruction automatically for writes and searches, so developers don’t need to add model-specific prompts to their application code.

General search is the default task. A field can instead be configured for classification, clustering or sentence_similarity. Individual searches can specify question_answering, fact_verification or code_retrieval_query. These choices let teams tune the embedding task to the operation without changing the basic flow of sending text to Pinecone. The supported task types are documented by Pinecone.

Google’s model supports text in more than 100 languages. Requests can be handled in any of Google’s regions, but no specific region is guaranteed. Teams with location requirements should account for that limit when evaluating the service.

Where the integration is available

Gemini Embedding 2 is supported on serverless indexes created with the Documents API. BYOC, pod-based and Records API indexes aren’t supported. Existing vector indexes can be converted to Documents API indexes without re-embedding, but adding integrated embedding to a converted index is not yet available. For now, teams need a new Documents API index to use this integration.

The Pinecone console and Python SDK version 10.3.0 support the model. Support for the Node.js and Go SDKs is planned, but isn’t available yet. The integrated embedding documentation explains how to configure a model on a Documents API index.

Although Gemini Embedding 2 can embed images, video and audio, Pinecone currently exposes it through this integration only for text fields. Multimodal support is planned for a later release. The practical fit today is therefore text search: teams can simplify embedding operations and combine vector retrieval with keyword search, provided they’re using a supported index type and can work within the plan’s token allowance.