langchain-google-genai 4.0.0 Released! (Consolidated SDK, Vertex deprecations, and more)

[Mason Daugherty (mdrxy)]

We are announcing a significant architectural overhaul for LangChain's Google AI integrations.

Our goal is to provide a unified, modern, and feature-complete experience while maintaining backward compatibility and smooth migration paths for existing users.

Unified Google GenAI SDK adoption

We've migrated langchain-google-genai to use the unified google-genai SDK, replacing the legacy google-ai-generativelanguage library.

As part of this, we will deprecate overlapping functionality in langchain-google-vertexai where the unified SDK provides equivalent capabilities.

Deprecations

• ChatVertexAI: Full feature parity with ChatGoogleGenerativeAI
• VertexAI: Full feature parity with GoogleGenerativeAI
• VertexAIEmbeddings: Full feature parity with GoogleGenerativeAIEmbeddings

Modules removed

• GoogleVectorStore: Deprecated service

Modules remaining in VertexAI:

• Vertex AI Platform Services
  • Model Garden integrations (VertexAIModelGarden)
  • MaaS (Model-as-a-Service) offerings (VertexModelGardenLlama, VertexModelGardenMistral)
  • Anthropic on Vertex AI (ChatAnthropicVertex)
  • Enterprise evaluation services (VertexStringEvaluator, VertexPairWiseStringEvaluator)
• Vertex AI Infrastructure:
  • Vector Search (Matching Engine) (VectorSearchVectorStore*)
  • Document storage integrations (GCSDocumentStorage, DataStoreDocumentStorage)
  • Vision models with specialized enterprise features (VertexAIImageCaptioning, VertexAIImageCaptioningChat, VertexAIVisualQnAChat, VertexAIImageEditorChat, VertexAIImageGeneratorChat)

As parity for these featured are made in the google-genai SDK, we will work to support these in langchain-google-genai (and follow suit with deprecating these in langchain-google-vertexai).

Impact assessment

Zero impact users (Majority)

• Who: Users of langchain-google-genai with standard configurations
• Impact: Automatic benefits from SDK migration with no code changes required
• Action: None - enjoy improved performance and reliability

Low impact users

• Who: Users of Gemini models via langchain-google-vertexai
• Impact: Deprecation warnings in logs, eventual migration needed. Will not be able to access the latest features supported in langchain-google-genai
• Action: Plan migration to langchain-google-genai for Gemini use cases

Medium impact users

• Who: Advanced users with custom transport configurations or low-level API access in langchain-google-genai

No impact users

• Who: Enterprise users leveraging only Vertex AI platform features (Model Garden, Vector Search, etc.)
• Impact: Continued security and critical support needs
• Action: Continue using langchain-google-vertexai for platform-specific needs

Breaking changes

Deprecated features

• langchain-google-genai
  • Drops gRPC transport options. REST is all that is available in the new SDK. If gRPC becomes available upstream, we will support it.
  • with_structured_output now defaults to method="json_schema" by default if not set. This makes use of the new native structured outputs
• langchain-google-vertexai
  • ChatVertexAI (use ChatGoogleGenerativeAI with the Vertex backend)
  • VertexAIEmbeddings (use GoogleGenerativeAIEmbeddings with the Vertex backend)

Structured output

• with_structured_output() now defaults to method="json_schema" instead of method="function_calling". This uses native structured outputs.

  # Before (implicit function_calling)
  llm.with_structured_output(MySchema)
  
  # After (implicit json_schema)
  llm.with_structured_output(MySchema)
  
  # To use the old behavior explicitly:
  llm.with_structured_output(MySchema, method="function_calling")

Removed features

• GoogleVectorStore: Removed. The upstream service has been deprecated by Google.

Migration guide

Chat Models: ChatVertexAI → ChatGoogleGenerativeAI

Also, for anyone still using legacy LLMs, VertexAI → GoogleGenerativeAI

Before (deprecated)

from langchain_google_vertexai import ChatVertexAI

llm = ChatVertexAI(
    model="gemini-3-pro-preview",
    temperature=0,
    max_tokens=1024,
    project="my-project",
    location="us-central1",
)

response = llm.invoke("Hello, world!")

After

Option 1: Environment variables (recommended)

export GOOGLE_GENAI_USE_VERTEXAI=true
export GOOGLE_CLOUD_PROJECT=my-project
export GOOGLE_CLOUD_LOCATION=us-central1  # Optional, defaults to us-central1

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-3-pro-preview",
    temperature=0,
    max_tokens=1024,
)

response = llm.invoke("Hello, world!")

Option 2: Explicit parameters

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-3-pro-preview",
    temperature=0,
    max_tokens=1024,
    project="my-project",
    location="us-central1",
    vertexai=True,  # Explicitly use Vertex AI
)

response = llm.invoke("Hello, world!")

Option 3: Vertex AI with API key

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-3-pro-preview",
    api_key="your-api-key",
    project="my-project",
    vertexai=True,
)

Embeddings: VertexAIEmbeddings → GoogleGenerativeAIEmbeddings

Before (deprecated)

from langchain_google_vertexai import VertexAIEmbeddings

embeddings = VertexAIEmbeddings(
    model_name="gemini-embedding-001",
    project="my-project",
    location="us-central1",
)

vectors = embeddings.embed_documents(["Hello", "World"])

After

from langchain_google_genai import GoogleGenerativeAIEmbeddings

embeddings = GoogleGenerativeAIEmbeddings(
    model="gemini-embedding-001",
    project="my-project",
    location="us-central1",
    vertexai=True,
)

vectors = embeddings.embed_documents(["Hello", "World"])

---

Backend auto-detection

When vertexai is not explicitly set, the backend is determined automatically:

1. If GOOGLE_GENAI_USE_VERTEXAI env var is set → uses that value
2. If credentials parameter is provided → uses Vertex AI
3. If project parameter is provided → uses Vertex AI
4. Otherwise → uses Gemini Developer API

Authentication

Backend	Authentication Methods
Gemini Developer API	API key via api_key param or GOOGLE_API_KEY/GEMINI_API_KEY env var
Vertex AI	API key, service account credentials, or Application Default Credentials (ADC)

Using Service Account Credentials

from google.oauth2 import service_account
from langchain_google_genai import ChatGoogleGenerativeAI

credentials = service_account.Credentials.from_service_account_file(
    "path/to/service-account.json",
    scopes=["https://www.googleapis.com/auth/cloud-platform"],
)

llm = ChatGoogleGenerativeAI(
    model="gemini-3-pro-preview",
    credentials=credentials,
    project="my-project-id",
)

Using Application Default Credentials (ADC)

gcloud auth application-default login

from langchain_google_genai import ChatGoogleGenerativeAI

# ADC is used automatically when no api_key is provided
llm = ChatGoogleGenerativeAI(
    model="gemini-3-pro-preview",
    project="my-project",
)

---

Environment variables

Variable	Purpose	Backend
GOOGLE_API_KEY	API key (primary)	Both
GEMINI_API_KEY	API key (fallback)	Both
GOOGLE_GENAI_USE_VERTEXAI	Force Vertex AI backend (true/false)	Vertex AI
GOOGLE_CLOUD_PROJECT	GCP project ID	Vertex AI
GOOGLE_CLOUD_LOCATION	GCP region (default: us-central1)	Vertex AI
HTTPS_PROXY	HTTP/HTTPS proxy URL	Both
SSL_CERT_FILE	Custom SSL certificate file	Both

---

Additional resources

For more resources, refer to the provider and reference docs.

[alejandroem1]

Hello, is this going to have an impact on the ChatAnthropicVertex class or will it continue to have future releases?

[ScsiOverdrive]

Also curious about this, as built-in tools don't seem to be supported by ChatAnthropicVertex yet

[Mason Daugherty (mdrxy)]

No impact. ChatAnthropicVertex will continue to get future releases but there may be some refactoring of the class to reduce the amount of work needed to port equivalent updates from ChatAnthropic. Though, I will admit ChatAnthropicVertex is lower on our priorities, so we welcome PRs ;)

[joseph (JCourt1)]

After upgrading from langchain-google-vertexai==3.0.3 to langchain_google_genai==4.0, and switching from ChatVertexAI to ChatGoogleGenerativeAI , we saw some strange latency issues.

We’re using gemini-2.5-flash-lite, with provisioned throughput.

We initially presumed that these latency issues were due to the fact that we had switched from structured output via “function_calling”, to structured output via “json_schema”, since this would change things to use “controlled generation” (which could add latency).

However, when we explicitly changed the code to pass method=”function_calling”, the latency issues remained.

Trying to dig into the issue, I believe that previously in langchain-google-vertexai, gRPC was used, whereas now in langchain_google_genai, REST is used. And on the whole, all of our llm calls seemed to get a bit slower, by a relatively small amount, ie. a couple of hundred ms. I presume this part of the issue may be due to the switch from grpc to rest?

But more significantly, one of our key llm calls in particular went from being consistently about 1 second, to consistently around 4 seconds (again, even when “function_calling” was explicitly passed, to replicate our pre-existing set-up). I assume the switch in protocol doesn’t account for that.

Nothing significant about the parameters that are sent seem to have changed. Another possible avenue for solving this - I assumed that maybe something about the class had changed, such that, for example, it wasn’t actually using vertex, or it wasn’t sending to our provisioned throughput. But we’re passing vertexai=True, and still passing the same region, and still setting the headers that we previously had:

additional_headers["X-Vertex-AI-LLM-Request-Type"] = "dedicated"

And indeed, checking in the vertex console in “Provisioned Throughput details”, we were recording traffic at the time this change was live (we’ve reverted it now).

So I’m pretty much stumped at this point. Does anyone have any idea what could have caused this discrepancy?

[Mason Daugherty (mdrxy)]

Hi joseph (@JCourt1), thanks for reporting. I suspect it may be related to this. However, in this case you are reporting a couple of hundred ms, so I'm not certain, as the PR I've linked addresses a much larger delay.

We'll keep an eye on things; would love to verify that this is not due to our implementation, but upstream.

[Angélica Viana (angelica-ti)]

Hi, Mason Daugherty (@mdrxy)! When will a version of langchain_community and langchain_google_firestore compatible with the langchain_google_genai package be released? I ask because I'm trying to use these packages together and I'm encountering dependency conflicts in langchain_core. The following error appears:

from langchain_core.memory import BaseMemory
ModuleNotFoundError: No module named 'langchain_core.memory'

[Mason Daugherty (mdrxy)]

Hi Angélica Viana (@angelica-ti), I believe langchain-community should currently be compatible. Would you mind sharing your pyproject.toml or a resolution error to help me understand the conflict?

Regarding langchain-google-firestore, that is not maintained by our team so I would encourage opening an issue in their repo.

[Angélica Viana (angelica-ti)]

The code threw an error at this exact line: from langchain.agents import AgentExecutor, create_tool_calling_agent, because the langchain-google-community and langchain-google-genai packages use different versions of langchain_core as a dependency, both in their latest versions (3.0.4 and 4.2.0 respectively).

[Mason Daugherty (mdrxy)]

Angélica Viana (@angelica-ti) can you confirm if you are referring to langchain-community or langchain-google-community? Your comments mention both

[Angélica Viana (angelica-ti)]

Sorry, It was my mistake! langchain-google-community.

[Aaron Lim (aaron-lim-es)]

Hi Mason Daugherty (@mdrxy)

Following the migration guidance in this announcement, we migrated from ChatVertexAI to ChatGoogleGenerativeAI with vertexai=True and ran A/B testing on our LangGraph agent using Gemini 2.5 Flash, and have been facing latency issues.

Setup:

• Gemini 2.5 Flash on Vertex AI (Cloud Run with ADC)
• langchain-google-genai 4.2.0
• Same model, prompts, and test suite for both experiments

Results - Median latency comparison:

Category	ChatVertexAI	ChatGoogleGenerativeAI	Regression
category 1 (no tool calling)	1.7s	5.9s	+247%
category 2 (with tool calling)	4.1s	7.7s	+88%
category 3 (with tool calling)	8.9s	15.6s	+75%
category 4 (with tool calling)	12.6s	21.0s	+67%
category 5 (with tool calling)	11.0s	20.5s	+86%

We're seeing 50-90% latency increases consistently across query types.
Is this expected due to REST vs gRPC overhead, or could there be other factors (connection reuse, serialization, etc.) contributing to this?

Also, are there any steps that we can take or others have done to mitigate this?
Or is the only way to go is to wait for this to be fixed in ChatGoogleGenerativeAI?

Any guidance on whether performance parity with ChatVertexAI is on the roadmap would be helpful.

[ccasadei]

I've experienced significant latency as well, specifically when using a Service Account (instead of an API Key). It appears to be related to authentication and token management/refreshing.

In my case, the first call to the embedder would take around 10 seconds. If I made another call immediately after, it took a fraction of a second. However, if I waited just one minute, the subsequent call would spike back up to 10 seconds.

I managed to solve this by manually updating the underlying libraries: google-auth==2.48.0 and google-genai==1.63.0.
Since these were being pulled in as transitive dependencies by other libraries, I didn't initially have control over their versions, and they were indeed outdated.

After the update, the initial cold start and "refresh" calls have dropped to around 1 second (or slightly more). While this is still a bit slower than using a raw API Key, it is much better and acceptable for my use case, as I prefer the security of using a Service Account.