Google AI chat models

Access Google AI's gemini and gemini-vision models, as well as other generative models through ChatGoogleGenerativeAI class in the langchain-google-genai integration package.

%pip install --upgrade --quiet  langchain-google-genai pillow

import getpass
import os

if "GOOGLE_API_KEY" not in os.environ:
    os.environ["GOOGLE_API_KEY"] = getpass.getpass("Provide your Google API Key")

Example usage

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(model="gemini-pro")
result = llm.invoke("Write a ballad about LangChain")
print(result.content)

In realms where data streams like fervent tides,
Where algorithms dance and knowledge abides,
A tale unfolds of LangChain, grand and bold,
A ballad sung in bits and bytes untold.

Amidst the codes and circuits' hum,
A spark ignited, a vision would come.
From minds of brilliance, a tapestry formed,
A model to learn, to comprehend, to transform.

In layers deep, its architecture wove,
A neural network, ever-growing, in love.
With language's essence, it sought to entwine,
To unlock the secrets, each word's design.

From texts vast and varied, it feasted and learned,
Its grasp on meaning, swiftly discerned.
Context and syntax, it embraced with grace,
Unraveling stories, at an astonishing pace.

Translations sprang forth, with seamless art,
Bridging tongues and weaving hearts apart.
From English to French, Chinese to Spanish, behold,
LangChain's prowess, like a language untold.

It summarized texts, with insights profound,
Extracting knowledge, without a sound.
Questions it answered, with eloquence rare,
A digital sage, beyond compare.

Yet, its journey faced trials and tribulations,
Obstacles that tested its dedication.
Biases and errors, it sought to transcend,
For fairness and accuracy, it would contend.

With every challenge, it emerged more strong,
Adaptive and resilient, all along.
For LangChain's purpose was etched in its core,
To empower humans, forevermore.

In classrooms and workplaces, it lent its hand,
A tireless assistant, at every demand.
It aided students, in their quests for knowledge,
And professionals thrived, with its guidance and homage.

As years unfurled, its fame grew wide,
A testament to its unwavering stride.
Researchers and scholars, they all took heed,
Of LangChain's brilliance, a groundbreaking deed.

And so, the ballad of LangChain resounds,
A tribute to progress, where innovation abounds.
In the annals of AI, its name shall be etched,
A pioneer, forever in our hearts sketched.

Gemini doesn't support SystemMessage at the moment, but it can be added to the first human message in the row. If you want such behavior, just set the convert_system_message_to_human to True:

from langchain_core.messages import HumanMessage, SystemMessage

model = ChatGoogleGenerativeAI(model="gemini-pro", convert_system_message_to_human=True)
model(
    [
        SystemMessage(content="Answer only yes or no."),
        HumanMessage(content="Is apple a fruit?"),
    ]
)

Streaming and Batching

ChatGoogleGenerativeAI natively supports streaming and batching. Below is an example.

for chunk in llm.stream("Write a limerick about LLMs."):
    print(chunk.content)
    print("---")
# Note that each chunk may contain more than one "token"

There once was an AI named Bert,
Whose language skills were quite expert.
---

With a vast dataset,
It could chat, even bet,
And write limericks, for what it's worth.
---

results = llm.batch(
    [
        "What's 2+2?",
        "What's 3+5?",
    ]
)
for res in results:
    print(res.content)

4
8

Multimodal support

To provide an image, pass a human message with contents of type List[dict], where each dict contains either an image value (type of image_url) or a text (type of text) value. The value of image_url can be any of the following:

• A public image URL
• An accessible gcs file (e.g., "gcs://path/to/file.png")
• A local file path
• A base64 encoded image (e.g., data:image/png;base64,abcd124)
• A PIL image

import requests
from IPython.display import Image

image_url = "https://picsum.photos/seed/picsum/300/300"
content = requests.get(image_url).content
Image(content)

from langchain_core.messages import HumanMessage
from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(model="gemini-pro-vision")
# example
message = HumanMessage(
    content=[
        {
            "type": "text",
            "text": "What's in this image?",
        },  # You can optionally provide text parts
        {"type": "image_url", "image_url": image_url},
    ]
)
llm.invoke([message])

AIMessage(content=' The image contains a snow-capped mountain peak.')

Gemini Prompting FAQs

As of the time this doc was written (2023/12/12), Gemini has some restrictions on the types and structure of prompts it accepts. Specifically:

1. When providing multimodal (image) inputs, you are restricted to at most 1 message of "human" (user) type. You cannot pass multiple messages (though the single human message may have multiple content entries)
2. System messages are not accepted.
3. For regular chat conversations, messages must follow the human/ai/human/ai alternating pattern. You may not provide 2 AI or human messages in sequence.
4. Message may be blocked if they violate the safety checks of the LLM. In this case, the model will return an empty response.

Safety Settings

Gemini models have default safety settings that can be overridden. If you are receiving lots of "Safety Warnings" from your models, you can try tweaking the safety_settings attribute of the model. For example, to turn off safety blocking for dangerous content, you can construct your LLM as follows:

from langchain_google_genai import (
    ChatGoogleGenerativeAI,
    HarmBlockThreshold,
    HarmCategory,
)

llm = ChatGoogleGenerativeAI(
    model="gemini-pro",
    safety_settings={
        HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE,
    },
)

For an enumeration of the categories and thresholds available, see Google's safety setting types.

Additional Configuration

You can pass the following parameters to ChatGoogleGenerativeAI in order to customize the SDK's behavior:

• client_options: Client Options to pass to the Google API Client, such as a custom client_options["api_endpoint"]
• transport: The transport method to use, such as rest, grpc, or grpc_asyncio.