Vertex AI¶

注意： Vertex 已基本被 Google GenAI 取代，后者通过 google-genai 包支持与 Vertex 相同的功能。请访问 Google GenAI 页面 获取最新示例和文档。

安装 Vertex AI¶

安装 Vertex AI 需要遵循以下步骤：

• 安装 Vertex Cloud SDK (https://googleapis.dev/python/aiplatform/latest/index.html)
• 设置默认项目、凭据和区域

服务账号基础认证示例¶

In [ ]:

%pip install llama-index-llms-vertex

%pip install llama-index-llms-vertex

In [ ]:

from llama_index.llms.vertex import Vertex
from google.oauth2 import service_account

filename = "vertex-407108-37495ce6c303.json"
credentials: service_account.Credentials = (
    service_account.Credentials.from_service_account_file(filename)
)
Vertex(
    model="text-bison", project=credentials.project_id, credentials=credentials
)

from llama_index.llms.vertex import Vertex from google.oauth2 import service_account filename = "vertex-407108-37495ce6c303.json" credentials: service_account.Credentials = ( service_account.Credentials.from_service_account_file(filename) ) Vertex( model="text-bison", project=credentials.project_id, credentials=credentials )

基础用法¶

调用 text-bison 模型的基本方法

In [ ]:

from llama_index.llms.vertex import Vertex
from llama_index.core.llms import ChatMessage, MessageRole

llm = Vertex(model="text-bison", temperature=0, additional_kwargs={})
llm.complete("Hello this is a sample text").text

from llama_index.llms.vertex import Vertex from llama_index.core.llms import ChatMessage, MessageRole llm = Vertex(model="text-bison", temperature=0, additional_kwargs={}) llm.complete("Hello this is a sample text").text

Out[ ]:

' ```\nHello this is a sample text\n```'

异步用法¶

异步¶

In [ ]:

(await llm.acomplete("hello")).text

(await llm.acomplete("hello")).text

Out[ ]:

' Hello! How can I help you?'

流式使用方式¶

流式传输¶

In [ ]:

list(llm.stream_complete("hello"))[-1].text

list(llm.stream_complete("hello"))[-1].text

Out[ ]:

' Hello! How can I help you?'

聊天使用指南¶

聊天生成¶

In [ ]:

chat = Vertex(model="chat-bison")
messages = [
    ChatMessage(role=MessageRole.SYSTEM, content="Reply everything in french"),
    ChatMessage(role=MessageRole.USER, content="Hello"),
]

chat = Vertex(model="chat-bison") messages = [ ChatMessage(role=MessageRole.SYSTEM, content="Reply everything in french"), ChatMessage(role=MessageRole.USER, content="Hello"), ]

In [ ]:

chat.chat(messages=messages).message.content

chat.chat(messages=messages).message.content

Out[ ]:

' Bonjour! Comment vas-tu?'

异步聊天¶

异步聊天响应¶

In [ ]:

(await chat.achat(messages=messages)).message.content

(await chat.achat(messages=messages)).message.content

Out[ ]:

' Bonjour! Comment vas-tu?'

流式聊天¶

流式聊天响应¶

In [ ]:

list(chat.stream_chat(messages=messages))[-1].message.content

list(chat.stream_chat(messages=messages))[-1].message.content

Out[ ]:

' Bonjour! Comment vas-tu?'

Gemini 模型¶

通过 Vertex AI 调用 Google Gemini 模型已获得完整支持。

Gemini Pro¶

In [ ]:

llm = Vertex(
    model="gemini-pro",
    project=credentials.project_id,
    credentials=credentials,
    context_window=100000,
)
llm.complete("Hello Gemini").text

llm = Vertex( model="gemini-pro", project=credentials.project_id, credentials=credentials, context_window=100000, ) llm.complete("Hello Gemini").text

Gemini 视觉模型¶

Gemini 视觉模型现支持通过 TextBlock 和 ImageBlock 处理结构化多模态输入，取代了旧有的基于字典的 content 格式。使用 blocks 参数可通过文件路径或 URL 嵌入文本与图像内容。

示例包含图片路径：

In [ ]:

from llama_index.llms.vertex import Vertex
from llama_index.core.llms import ChatMessage, TextBlock, ImageBlock

history = [
    ChatMessage(
        role="user",
        blocks=[
            ImageBlock(path="sample.jpg"),
            TextBlock(text="What is in this image?"),
        ],
    ),
]
llm = Vertex(
    model="gemini-1.5-flash",
    project=credentials.project_id,
    credentials=credentials,
    context_window=100000,
)
print(llm.chat(history).message.content)

from llama_index.llms.vertex import Vertex from llama_index.core.llms import ChatMessage, TextBlock, ImageBlock history = [ ChatMessage( role="user", blocks=[ ImageBlock(path="sample.jpg"), TextBlock(text="What is in this image?"), ], ), ] llm = Vertex( model="gemini-1.5-flash", project=credentials.project_id, credentials=credentials, context_window=100000, ) print(llm.chat(history).message.content)

The image shows a man in handcuffs being escorted by police officers. The man is wearing a black jacket and a black hooded sweatshirt. He is being escorted by two police officers who are wearing black balaclavas and bulletproof vests. The officers are from the Romanian Gendarmerie. The image is likely from a news report or a documentary about a crime.

示例包含图片 URL：

In [ ]:

from llama_index.llms.vertex import Vertex
from llama_index.core.llms import ChatMessage, TextBlock, ImageBlock

history = [
    ChatMessage(
        role="user",
        blocks=[
            ImageBlock(
                url="https://upload.wikimedia.org/wikipedia/commons/7/71/Sibirischer_tiger_de_edit02.jpg"
            ),
            TextBlock(text="What is in this image?"),
        ],
    ),
]
llm = Vertex(
    model="gemini-1.5-flash",
    project=credentials.project_id,
    credentials=credentials,
    context_window=100000,
)
print(llm.chat(history).message.content)

from llama_index.llms.vertex import Vertex from llama_index.core.llms import ChatMessage, TextBlock, ImageBlock history = [ ChatMessage( role="user", blocks=[ ImageBlock( url="https://upload.wikimedia.org/wikipedia/commons/7/71/Sibirischer_tiger_de_edit02.jpg" ), TextBlock(text="What is in this image?"), ], ), ] llm = Vertex( model="gemini-1.5-flash", project=credentials.project_id, credentials=credentials, context_window=100000, ) print(llm.chat(history).message.content)

A close-up of a tiger's face. The tiger is looking directly at the camera with a serious expression. The fur is a mix of orange, black, and white. The background is blurred, making the tiger the main focus of the image.