SunsetDiff

Vertex AI text-embedding-004, text-embedding-005, and text-multilingual-embedding-002 shutdown on April 1, 2027

The dataset has 4 Google Vertex ids with shutdown date 2027-04-01: multimodalembedding@001, text-embedding-004, text-embedding-005, text-multilingual-embedding-002. The embeddings table on the model versions page lists retirement date April 1, 2027 for each of them. The Replacement model ID cell on each of those rows is empty. The replacements field is empty, so each model page says Not published. The page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed."

Last reviewed . Publisher: SunsetDiff.

Steps

  1. Find the four ids. Search code, config, and vector index settings for multimodalembedding@001, text-embedding-004, text-embedding-005, and text-multilingual-embedding-002. A hit is a candidate. Each of those ids is a SunsetDiff row. The model versions page lists retirement date April 1, 2027 for each of them.
  2. Read the empty replacement cell. The Replacement model ID cell is empty for multimodalembedding@001, text-embedding-004, text-embedding-005, text-multilingual-embedding-002. The get text embeddings page describes gemini-embedding-001 as a model that unifies text-embedding-005 and text-multilingual-embedding-002. That sentence is a description. It is not a filled Replacement model ID cell, and it does not name text-embedding-004. Pick the model id you will call, and write it down. This guide does not fill the cell Google left empty.
  3. Re-embed or build a second index. A new embedding model writes a new list of numbers. Those numbers do not drop into an index built from the old model. Embed the corpus again with the model you picked, or keep the old index and search a second index until the new one is ready. Check the vector length and the task_type on both the documents and the queries.
  4. Test one query before the date. Search the new index with one query you use now, embedded by the same model, before April 1, 2027. A successful call to the old id does not test the new index.

Models and dates

These 4 ids are the Google Vertex rows whose shutdown date is 2027-04-01. The model id links to its SunsetDiff model page. Release date is the date stored from the model versions page. The page prints February 12, 2024 for multimodalembedding@001, May 14, 2024 for text-embedding-004, November 18, 2024 for text-embedding-005, and May 14, 2024 for text-multilingual-embedding-002. Shutdown date is the retirement date. Replacement model ID is the cell on that page. Not published means the cell is empty and the stored replacements field is empty. These rows do not store a deprecation_date.

The model versions page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed." Checked on 2026-11-04, the embeddings table lists retirement date April 1, 2027 on each of these 4 rows. multimodalembedding@001 is one of those rows. Its release date on the page is February 12, 2024. The Replacement model ID cell is empty.

4 Vertex embedding ids with retirement date 2027-04-01.
Model idRelease dateShutdown dateReplacement model ID
multimodalembedding@0012024-02-122027-04-01Not published
text-embedding-0042024-05-142027-04-01Not published
text-embedding-0052024-11-182027-04-01Not published
text-multilingual-embedding-0022024-05-142027-04-01Not published

Source: Vertex AI model versions , checked November 4, 2026.

The replacement cell is empty

The replacements field is empty on multimodalembedding@001, text-embedding-004, text-embedding-005, text-multilingual-embedding-002. The model page under Replacements then says Not published. The model versions page also leaves the Replacement model ID cell empty.

The get text embeddings page says this about gemini-embedding-001: "State-of-the-art performance across English, multilingual and code tasks. It unifies the previously specialized models like text-embedding-005 and text-multilingual-embedding-002 and achieves better performance in their respective domains." That sentence names text-embedding-005 and text-multilingual-embedding-002. It does not name text-embedding-004. The same sentence is on the text embeddings API page and on the Agent Retrieval supported-models table. None of those pages put gemini-embedding-001 in the Replacement model ID cell for the April 1 rows.

gemini-embedding-2 is on the same embeddings table. Release date 2026-04-22. The retirement cell is empty. The stored shutdown kind is none. See gemini-embedding-2.

gemini-embedding-001 is on that table with release date May 20, 2025 and retirement date No sooner than May 20, 2028. The stored shutdown date is 2028-05-20 and the stored kind is not_before. The replacement cell is empty. See gemini-embedding-001.

On the retired models table, the recommended upgrade for textembedding-gecko@003, textembedding-gecko-multilingual@001, textembedding-gecko@002, and textembedding-gecko@001 is gemini-embedding-001. That column is Recommended upgrade, and those rows are already retired. It is a different table from the embeddings models table.

The multimodal embeddings API page lists gemini-embedding-2 in the supported models table, in the Code column, for Embeddings for Multimodal. The model versions page leaves the Replacement model ID cell empty for multimodalembedding@001. This guide does not copy gemini-embedding-2 into that cell. Code samples on the multimodal embeddings API page still send multimodalembedding@001.

Other rows on the embeddings table of the model versions page.
Model idRelease date on the pageRetirement date on the pageReplacement model ID
gemini-embedding-2April 22, 2026EmptyNot published
gemini-embedding-001May 20, 2025No sooner than May 20, 2028Not published

Sources: Vertex AI model versions , Multimodal embeddings API , checked November 4, 2026.

A new model writes different vectors

An embedding model turns text into a list of numbers. Search compares those lists. If you store lists from one model and then ask a second model to embed the query, the numbers are not a pair. You embed the stored text again, or you keep two indexes during the move: the old index for the old model, and a new index for the new model.

The Spanner backfill page says: "This is because the embedding vectors generated for the same piece of text might differ depending on the version of the model that you use; which is why you want to avoid using different model versions to generate embeddings in the same dataset." Then: "In addition, updating the model version in the model definition statement doesn't update the embeddings that are already generated with this model." Changing the model string on the next request leaves the stored lists as they are.

The RAG Engine page says: "The association between your embedding model and the RAG corpus remains fixed for the lifetime of your RAG corpus." Under the heading When Gecko models are discontinued, it says: "When your Gecko model is discontinued, you must migrate the RAG corpus, which means that you create a new RAG corpus and re-import the data." That paragraph names Gecko models. It does not name text-embedding-004, text-embedding-005, or text-multilingual-embedding-002.

The get text embeddings page says: "All models produce a full-length embedding vector by default. For gemini-embedding-001, this vector has 3072 dimensions, and other models produce 768-dimensional vectors." Then: "However, by using the output_dimensionality parameter, users can control the size of the output embedding vector." The text embeddings API says outputDimensionality is: "Used to specify output embedding size. If set, output embeddings will be truncated to the size specified." The Python sample on the get text embeddings page calls that field output_dimensionality. Set the length your index expects. The Spanner page says the same text can still produce different numbers when the model version changes, so embed the corpus with the one model you will keep.

Sources: Spanner backfill embeddings , Use embedding models with RAG Engine , Get text embeddings , Text embeddings API , checked November 4, 2026.

Output size and task type

The Agent Retrieval page lists max output dimensions 768 for text-embedding-004, text-embedding-005, and text-multilingual-embedding-002. It lists 3072 for gemini-embedding-001 and for gemini-embedding-2. It lists max sequence length 2048 tokens for the three retiring text models and for gemini-embedding-001, and 8192 tokens for gemini-embedding-2. It says text-embedding-004 and text-embedding-005 are English, and text-multilingual-embedding-002 is for multilingual tasks.

The get text embeddings supported-models table says up to 768 for text-embedding-005 and text-multilingual-embedding-002, and up to 3072 for gemini-embedding-001 and gemini-embedding-2. Checked on the page that says Last updated 2026-11-03 UTC, that table does not list text-embedding-004. The model versions page and the Agent Retrieval table still list text-embedding-004.

The text embeddings API says task_type: "If left blank, the default used is RETRIEVAL_QUERY." It says title is: "Only valid with task_type=RETRIEVAL_DOCUMENT." The task types page says: "If your embeddings use case doesn't align with a documented use case, use the RETRIEVAL_QUERY task type by default." And: "To get embeddings that you can use for information retrieval, use the RETRIEVAL_DOCUMENT task type to embed your documents and the RETRIEVAL_QUERY task type to embed user queries." It also says: "text-embedding-005 supports a new task type CODE_RETRIEVAL_QUERY, which can be used to retrieve relevant code blocks using plain text queries." Code blocks on that page use RETRIEVAL_DOCUMENT. The query uses CODE_RETRIEVAL_QUERY.

The task types page says task types are supported by text-embedding-005, text-multilingual-embedding-002, and gemini-embedding-001. That list does not include text-embedding-004 or gemini-embedding-2. The Agent Retrieval sample still sets task_type to RETRIEVAL_DOCUMENT next to model_id text-embedding-004. Use the task type you used when you embedded the corpus. A query and a document need the pair the task types page names.

The Agent Retrieval page also says: "Support for multimodal embedding (multimodalembedding@001) is coming soon." The model versions page already lists multimodalembedding@001 with retirement date 2027-04-01. Those two sentences are on different pages. The Agent Retrieval dimensions table does not list multimodalembedding@001.

Sources: Get text embeddings , Text embeddings API , Choose an embeddings task type , Autogenerating embeddings for Agent Retrieval , checked November 4, 2026.

Change the model id

The text embeddings API syntax builds the client with TextEmbeddingModel.from_pretrained(MODEL_ID). MODEL_ID is the string you change. The page prints get_embeddings(...) with an ellipsis. This guide keeps that ellipsis.

The get text embeddings page tells you to set GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, and GOOGLE_GENAI_USE_ENTERPRISE=True, then call genai.Client() with no arguments. Its Python sample sets model to gemini-embedding-001 and passes task_type and output_dimensionality on EmbedContentConfig. The sample on the page sends three strings in contents. The copy below sends one string, in the same list. That sample is a call shape. The model versions page does not put gemini-embedding-001 in the Replacement model ID cell.

The predict URL on that page is https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID:predict. The sample sets the last path segment to gemini-embedding-001. The region in the sample is us-central1.

Search the repository before the date. A host console or a vector index can hold an id that git does not show. How to find every model ID in your codebase covers SDK argument names. The scan page builds a search from dataset ids that have a shutdown date. The search below is the four stored rows, including multimodalembedding@001.

Text embeddings API syntax. Set MODEL_ID to the id in the call you have now.

PROJECT_ID = PROJECT_ID
REGION = us-central1
MODEL_ID = MODEL_ID

import vertexai
from vertexai.language_models import TextEmbeddingModel

vertexai.init(project=PROJECT_ID, location=REGION)

model = TextEmbeddingModel.from_pretrained(MODEL_ID)
embeddings = model.get_embeddings(...)

Get text embeddings Python sample. The model string in the sample is gemini-embedding-001.

from google import genai
from google.genai.types import EmbedContentConfig

client = genai.Client()
response = client.models.embed_content(
    model="gemini-embedding-001",
    contents=[
        "How do I get a driver's license/learner's permit?",
    ],
    config=EmbedContentConfig(
        task_type="RETRIEVAL_DOCUMENT",
        output_dimensionality=3072,
        title="Driver's License",
    ),
)
print(response)

ripgrep. These four ids are the stored rows.

rg -n -w -e 'multimodalembedding@001' -e 'text-embedding-004' -e 'text-embedding-005' -e 'text-multilingual-embedding-002'

Sources: Text embeddings API , Get text embeddings , Autogenerating embeddings for Agent Retrieval , checked November 4, 2026.

Checklist before April 1, 2027

Do this before April 1, 2027. The model versions page says: "if you don't have time to follow the guide and just need to quickly resolve the errors caused by models reaching their retirement dates, do the following:" Then: "Update your application to point to the recommended upgrades." Then: "Test all mission critical features to make sure everything works as expected." Then: "Deploy the updates like you normally would." The replacement cell is empty, so that page does not name the upgrade string. The same page does not print an HTTP status for these embedding rows.

The Gemini API row for text-embedding-004 stores shutdown date 2026-01-14 and replacement gemini-embedding-2. Its source was last checked 2026-11-12. That row is the Gemini API. These 4 rows are Vertex AI. See Gemini API text-embedding-004. OpenAI's April 1, 2027 list is a third date, on GPT-5.1, gpt-5.3-codex, and gpt-5.4-nano shutdown on April 1, 2027.

Sources: Vertex AI model versions , Get text embeddings , Choose an embeddings task type , Gemini API deprecations , checked November 4, 2026.

Related SunsetDiff pages

All guides

Questions

Which Vertex embedding models retire on April 1, 2027?

The dataset lists multimodalembedding@001, text-embedding-004, text-embedding-005, text-multilingual-embedding-002 with shutdown date 2027-04-01. The model versions page lists retirement date April 1, 2027 for each of them, including multimodalembedding@001.

What is the official replacement?

The Replacement model ID cell is empty for multimodalembedding@001, text-embedding-004, text-embedding-005, text-multilingual-embedding-002, and the stored replacements field is empty. The model page says Not published. The get text embeddings page says gemini-embedding-001 unifies text-embedding-005 and text-multilingual-embedding-002. That sentence does not fill the replacement cell, and it does not name text-embedding-004.

Can I keep my stored vectors and only change the model id?

The stored vectors stay as they are. Changing the model id writes new vectors for new requests. The Spanner backfill page says the vectors for the same text might differ by model version, and that updating the model version does not update embeddings already generated. Embed the corpus again, or search a second index while the old index still uses the old model. The RAG Engine page says: "The association between your embedding model and the RAG corpus remains fixed for the lifetime of your RAG corpus."

What output size and task type should I check?

The Agent Retrieval table lists 768 dimensions for text-embedding-004, text-embedding-005, and text-multilingual-embedding-002, and 3072 for gemini-embedding-001. That table does not list multimodalembedding@001. outputDimensionality truncates the vector to the size you set. The text embeddings API says a blank task_type uses RETRIEVAL_QUERY. Documents use RETRIEVAL_DOCUMENT. The task types page lists task type support for text-embedding-005, text-multilingual-embedding-002, and gemini-embedding-001. It does not list text-embedding-004.

What does Google say happens on April 1, 2027?

The model versions page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed." It also says to update the application to the recommended upgrades, test mission critical features, and deploy. It does not print an HTTP status for these embedding rows, and the Replacement model ID cell is empty.

Is this the same date as the Gemini API text-embedding-004 row?

No. The Gemini API row stores shutdown date 2026-01-14 and replacement gemini-embedding-2. These Vertex rows store 2027-04-01. OpenAI's April 1, 2027 list is a third set of ids.

More guides

Deprecation alerts

We email you before models shut down. Unsubscribe anytime.

Deprecation alert signup
Providers (optional)