ChatGoogleGenerativeAI
Google AI offers a number of different chat models, including the powerful Gemini series. For information on the latest models, their features, context windows, etc. head to the Google AI docs.
This will help you getting started with ChatGoogleGenerativeAI
chat
models. For detailed documentation of all
ChatGoogleGenerativeAI
features and configurations head to the API
reference.
Overviewβ
Integration detailsβ
Class | Package | Local | Serializable | PY support | Package downloads | Package latest |
---|---|---|---|---|---|---|
ChatGoogleGenerativeAI | @langchain/google-genai | β | β | β |
Model featuresβ
See the links in the table headers below for guides on how to use specific features.
Tool calling | Structured output | JSON mode | Image input | Audio input | Video input | Token-level streaming | Token usage | Logprobs |
---|---|---|---|---|---|---|---|---|
β | β | β | β | β | β | β | β | β |
Setupβ
You can access Googleβs gemini
and gemini-vision
models, as well as
other generative models in LangChain through ChatGoogleGenerativeAI
class in the @langchain/google-genai
integration package.
You can also access Google's gemini
family of models via the LangChain VertexAI and VertexAI-web integrations.
Click here to read the docs.
Credentialsβ
Get an API key here: https://ai.google.dev/tutorials/setup
Then set the GOOGLE_API_KEY
environment variable:
export GOOGLE_API_KEY="your-api-key"
If you want to get automated tracing of your model calls you can also set your LangSmith API key by uncommenting below:
# export LANGCHAIN_TRACING_V2="true"
# export LANGCHAIN_API_KEY="your-api-key"
Installationβ
The LangChain ChatGoogleGenerativeAI
integration lives in the
@langchain/google-genai
package:
- npm
- yarn
- pnpm
npm i @langchain/google-genai
yarn add @langchain/google-genai
pnpm add @langchain/google-genai
Instantiationβ
Now we can instantiate our model object and generate chat completions:
import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
const llm = new ChatGoogleGenerativeAI({
model: "gemini-1.5-pro",
temperature: 0,
maxRetries: 2,
// other params...
});
Invocationβ
const aiMsg = await llm.invoke([
[
"system",
"You are a helpful assistant that translates English to French. Translate the user sentence.",
],
["human", "I love programming."],
]);
aiMsg;
AIMessage {
"content": "J'adore programmer. \n",
"additional_kwargs": {
"finishReason": "STOP",
"index": 0,
"safetyRatings": [
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HARASSMENT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"probability": "NEGLIGIBLE"
}
]
},
"response_metadata": {
"finishReason": "STOP",
"index": 0,
"safetyRatings": [
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HARASSMENT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"probability": "NEGLIGIBLE"
}
]
},
"tool_calls": [],
"invalid_tool_calls": [],
"usage_metadata": {
"input_tokens": 21,
"output_tokens": 5,
"total_tokens": 26
}
}
console.log(aiMsg.content);
J'adore programmer.
Chainingβ
We can chain our model with a prompt template like so:
import { ChatPromptTemplate } from "@langchain/core/prompts";
const prompt = ChatPromptTemplate.fromMessages([
[
"system",
"You are a helpful assistant that translates {input_language} to {output_language}.",
],
["human", "{input}"],
]);
const chain = prompt.pipe(llm);
await chain.invoke({
input_language: "English",
output_language: "German",
input: "I love programming.",
});
AIMessage {
"content": "Ich liebe das Programmieren. \n",
"additional_kwargs": {
"finishReason": "STOP",
"index": 0,
"safetyRatings": [
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HARASSMENT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"probability": "NEGLIGIBLE"
}
]
},
"response_metadata": {
"finishReason": "STOP",
"index": 0,
"safetyRatings": [
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_HARASSMENT",
"probability": "NEGLIGIBLE"
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"probability": "NEGLIGIBLE"
}
]
},
"tool_calls": [],
"invalid_tool_calls": [],
"usage_metadata": {
"input_tokens": 16,
"output_tokens": 7,
"total_tokens": 23
}
}
Safety Settingsβ
Gemini models have default safety settings that can be overridden. If
you are receiving lots of βSafety Warningsβ 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 import enums
from the @google/generative-ai
package, then construct your LLM as
follows:
import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
import { HarmBlockThreshold, HarmCategory } from "@google/generative-ai";
const llmWithSafetySettings = new ChatGoogleGenerativeAI({
model: "gemini-1.5-pro",
temperature: 0,
safetySettings: [
{
category: HarmCategory.HARM_CATEGORY_HARASSMENT,
threshold: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
},
],
// other params...
});
Tool callingβ
Tool calling with Google AI is mostly the same as tool calling with other models, but has a few restrictions on schema.
The Google AI API does not allow tool schemas to contain an object with unknown properties. For example, the following Zod schemas will throw an error:
const invalidSchema = z.object({ properties: z.record(z.unknown()) });
and
const invalidSchema2 = z.record(z.unknown());
Instead, you should explicitly define the properties of the object field. Hereβs an example:
import { tool } from "@langchain/core/tools";
import { ChatGoogleGenerativeAI } from "@langchain/google-genai";
import { z } from "zod";
// Define your tool
const fakeBrowserTool = tool(
(_) => {
return "The search result is xyz...";
},
{
name: "browser_tool",
description:
"Useful for when you need to find something on the web or summarize a webpage.",
schema: z.object({
url: z.string().describe("The URL of the webpage to search."),
query: z.string().optional().describe("An optional search query to use."),
}),
}
);
const llmWithTool = new ChatGoogleGenerativeAI({
model: "gemini-pro",
}).bindTools([fakeBrowserTool]); // Bind your tools to the model
const toolRes = await llmWithTool.invoke([
[
"human",
"Search the web and tell me what the weather will be like tonight in new york. use a popular weather website",
],
]);
console.log(toolRes.tool_calls);
[
{
name: 'browser_tool',
args: {
url: 'https://www.weather.com',
query: 'weather tonight in new york'
},
type: 'tool_call'
}
]
Gemini Prompting FAQsβ
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:
- When providing multimodal (image) inputs, you are restricted to at most 1 message of βhumanβ (user) type. You cannot pass multiple messages (though the single human message may have multiple content entries)
- System messages are not natively supported, and will be merged with the first human message if present.
- 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.
- Message may be blocked if they violate the safety checks of the LLM. In this case, the model will return an empty response.
API referenceβ
For detailed documentation of all ChatGoogleGenerativeAI features and configurations head to the API reference: https://api.js.langchain.com/classes/langchain_google_genai.ChatGoogleGenerativeAI.html
Relatedβ
- Chat model conceptual guide
- Chat model how-to guides