Gemini Embedding 2 (gemini-embedding-2)

Google’s Gemini Embedding 2 embeds text for the task you choose, such as search or clustering, and works with integrated embedding on document indexes.

​

When to use it

Gemini Embedding 2 (gemini-embedding-2) is a dense text embedding model from Google. Pinecone offers it for integrated embedding on document indexes, so you upsert and search with text and Pinecone embeds it through Google Vertex AI.

The model embeds text differently for each task you choose with task_type, such as retrieval_document for stored text, question_answering or fact_verification for queries, or clustering for grouping texts. You can also override the query task for one search.

​

Inputs, limits, and pricing

Input	Limits	Price
Text	8192 tokens per input, 96 inputs per request, 10,000 bytes per search query	$0.20 per 1M tokens

Monthly and per-minute limits depend on your plan. See Rate limits.

​

Quickstart

​

Installation

Integrated embedding on document indexes requires Python SDK 10.1.0, Node.js SDK 9.1.0, or Go SDK 7.1.0 or later.

Python

pip install --upgrade pinecone

​

Create index

This creates a document index whose chunk_text field embeds with gemini-embedding-2 at 768 dimensions. The field’s write task defaults to retrieval_document, and its queries default to retrieval_query.

Python

import time
from pinecone import Pinecone, SchemaBuilder

pc = Pinecone(api_key="YOUR_API_KEY")

schema = (
    SchemaBuilder()
      .add_string_field(
          name="chunk_text",
          embed={"model": "gemini-embedding-2", "dimension": 768},
      )
      .build()
)
pc.indexes.create(name="gemini-embedding-2", schema=schema)

while not pc.indexes.describe(name="gemini-embedding-2").status.ready:
    time.sleep(2)

index = pc.index(name="gemini-embedding-2")

​

Embed & upsert

Upsert documents with plain text on the embedded field. Pinecone embeds each value on write, up to 96 documents per request.

Python

index.documents.upsert(
    namespace="example-namespace",
    documents=[
        {"_id": "doc1", "chunk_text": "Apple is a popular fruit known for its sweetness and crisp texture."},
        {"_id": "doc2", "chunk_text": "The tech company Apple is known for its innovative products like the iPhone."},
        {"_id": "doc3", "chunk_text": "Many people enjoy eating apples as a healthy snack."},
        {"_id": "doc4", "chunk_text": "Apple Inc. has revolutionized the tech industry with its sleek designs and user-friendly interfaces."},
        {"_id": "doc5", "chunk_text": "An apple a day keeps the doctor away, as the saying goes."},
        {"_id": "doc6", "chunk_text": "Apple Computer Company was founded on April 1, 1976, by Steve Jobs, Steve Wozniak, and Ronald Wayne as a partnership."},
    ],
)

​

Query

Search with an embed clause, and Pinecone embeds the query text with the field’s read task.

Python

response = index.documents.search(
    namespace="example-namespace",
    top_k=3,
    score_by=[
        {"type": "embed", "fields": ["chunk_text"], "query": "Tell me about the tech company known as Apple."}
    ],
    include_fields=["chunk_text"],
)

for match in response.matches:
    print(match._id, match._score, match.chunk_text)

​

Parameters

Set these in the field’s write_parameters and read_parameters. Set dimension and processing_location on the embed object itself, as embed.dimension and embed.processing_location. Putting either one in write_parameters or read_parameters returns a 400.

Parameter	Type	Required/Optional	Description	Default
task_type	string	Optional	What the embedding is for. Accepted values: retrieval_document, retrieval_query, question_answering, fact_verification, code_retrieval_query, classification, clustering, or sentence_similarity. See Task types.	retrieval_document on write. On read, retrieval_query for a retrieval_document field, or the same task for a symmetric field.
truncate	string	Optional	How to handle inputs longer than those supported by the model. Accepted values: END or NONE. END truncates the input sequence at the input token limit. NONE returns an error when the input exceeds the input token limit.	END

​

Task types

Each task type is a write task, a query task, or a symmetric task. Set the write task in the field’s write_parameters and the default query task in its read_parameters. A symmetric task embeds stored text and queries the same way, for comparing texts with each other rather than searching. For which tasks a field accepts, see Set task types on the field.

For a field you search, the defaults fit general search. If a field’s queries are mostly questions answered by the stored text, claims to check against it, or plain-language lookups for code, set the matching query task in its read_parameters, or override the query task for one search. Use a symmetric task only for a field whose texts you compare with each other instead of searching, because you can’t change a field’s write task after you create the index.

Task type	Kind	Use it for	Prompt
retrieval_document	Write	Stored text that you search. This is the default write task.	title: none | text:
retrieval_query	Query	General search queries. This is the default query task for a retrieval_document field.	task: search result | query:
question_answering	Query	Questions whose answers are in the stored text.	task: question answering | query:
fact_verification	Query	Statements to check against the stored text.	task: fact checking | query:
code_retrieval_query	Query	Natural-language queries for stored code.	task: code retrieval | query:
classification	Symmetric	Texts you sort into labeled categories.	task: classification | query:
clustering	Symmetric	Texts you group by similarity.	task: clustering | query:
sentence_similarity	Symmetric	Texts you compare for semantic similarity.	task: sentence similarity | query:

The model reads the task from a short prompt at the start of each input, in the formats Google defines in Task types with Embeddings 2. Pinecone adds the prompt for you from the task type, so send plain text and don’t add a prompt yourself. retrieval_document uses Google’s document format, which has a title slot instead of a task, and Pinecone always sets the title to none. A symmetric task adds the same prompt to stored text and to queries. For example, a search for “Is the Eiffel Tower taller than 300 m?” on a field whose read_parameters.task_type is question_answering reaches the model as:

Question answering prompt

task: question answering | query: Is the Eiffel Tower taller than 300 m?

The prompt counts as input tokens. It adds about 6 or 7 tokens to each value and query, it’s included in embed_total_tokens, and it counts toward the 8,192-token limit, so with truncate: NONE the text itself can be at most about 8,185 tokens. A value truncated at the limit can report up to 8,256 tokens.

​

Override the query task

To embed one query with a different task, set parameters on the embed clause. The Python SDK supports parameters in 10.3.0 and later. The Node.js and Go SDKs don’t support it yet, so use the REST API with them. This search embeds its query as a question with question_answering:

Python

response = index.documents.search(
    namespace="example-namespace",
    top_k=3,
    score_by=[
        {
            "type": "embed",
            "fields": ["chunk_text"],
            "query": "Who founded Apple?",
            "parameters": {"task_type": "question_answering"},
        }
    ],
    include_fields=["chunk_text"],
)

​

Use a symmetric task

This creates an index that embeds support tickets with clustering. Its queries use clustering too, so a search returns the tickets most similar to the one you pass in, which you can use to find duplicates or group related issues.

Python

schema = (
    SchemaBuilder()
      .add_string_field(
          name="ticket_text",
          embed={
              "model": "gemini-embedding-2",
              "write_parameters": {"task_type": "clustering"},
          },
      )
      .build()
)
pc.indexes.create(name="support-tickets", schema=schema)

After you upsert tickets, search with a new ticket’s text to find similar ones. A search query is limited to 10,000 bytes, so shorten a longer ticket before you search with it:

Python

index = pc.index(name="support-tickets")

response = index.documents.search(
    namespace="example-namespace",
    top_k=5,
    score_by=[
        {"type": "embed", "fields": ["ticket_text"], "query": "Login fails after a password reset."}
    ],
    include_fields=["ticket_text"],
)

​

Good to know

• Dimension: 3,072 by default, or any integer from 128 to 3,072. The metric is cosine only.
• Input: Text only. Pinecone doesn’t support image, audio, or video input for this model yet.
• Processing location: global only. See Processing locations.
• Availability: Integrated embedding on document indexes only, on the Builder, Standard, and Enterprise plans.