使用谷歌Gemini模型实现多模态LLM图像理解及LlamaIndex检索增强生成¶

本笔记本演示如何利用谷歌Gemini Vision模型进行图像理解。

首先展示Gemini当前支持的几项功能：

• complete（同步/异步）：处理单个提示词和图片列表
• chat（同步/异步）：处理多轮对话消息
• stream complete（同步/异步）：实现complete功能的流式输出
• stream chat（同步/异步）：实现chat功能的流式输出

第二部分将尝试结合Gemini与Pydantic从谷歌地图图片中解析结构化信息：

• 定义包含属性字段的Pydantic类
• 让gemini-pro-vision模型理解每张图片并输出结构化结果

第三部分提出使用Gemini和LlamaIndex为小型谷歌地图餐厅数据集构建简易检索增强生成流程：

• 基于第二步的结构化输出构建向量索引
• 使用gemini-pro模型综合结果，根据用户查询推荐餐厅

注意：google-generativeai仅在特定国家和地区可用。

In [ ]:

%pip install llama-index-multi-modal-llms-gemini
%pip install llama-index-vector-stores-qdrant
%pip install llama-index-embeddings-gemini
%pip install llama-index-llms-gemini

%pip install llama-index-multi-modal-llms-gemini %pip install llama-index-vector-stores-qdrant %pip install llama-index-embeddings-gemini %pip install llama-index-llms-gemini

In [ ]:

!pip install llama-index 'google-generativeai>=0.3.0' matplotlib qdrant_client

!pip install llama-index 'google-generativeai>=0.3.0' matplotlib qdrant_client

使用 Gemini 理解来自 URL 的图像¶

In [ ]:

%env GOOGLE_API_KEY=...

%env GOOGLE_API_KEY=...

In [ ]:

import os

GOOGLE_API_KEY = ""  # add your GOOGLE API key here
os.environ["GOOGLE_API_KEY"] = GOOGLE_API_KEY

import os GOOGLE_API_KEY = "" # add your GOOGLE API key here os.environ["GOOGLE_API_KEY"] = GOOGLE_API_KEY

初始化 GeminiMultiModal 并从 URL 加载图像¶

In [ ]:

from llama_index.llms.gemini import Gemini
from llama_index.core.llms import ChatMessage, ImageBlock


image_urls = [
    "https://storage.googleapis.com/generativeai-downloads/data/scene.jpg",
    # Add yours here!
]
gemini_pro = Gemini(model_name="models/gemini-1.5-flash")
msg = ChatMessage("Identify the city where this photo was taken.")
for img_url in image_urls:
    msg.blocks.append(ImageBlock(url=img_url))

from llama_index.llms.gemini import Gemini from llama_index.core.llms import ChatMessage, ImageBlock image_urls = [ "https://storage.googleapis.com/generativeai-downloads/data/scene.jpg", # Add yours here! ] gemini_pro = Gemini(model_name="models/gemini-1.5-flash") msg = ChatMessage("Identify the city where this photo was taken.") for img_url in image_urls: msg.blocks.append(ImageBlock(url=img_url))

In [ ]:

from PIL import Image
import requests
from io import BytesIO
import matplotlib.pyplot as plt

img_response = requests.get(image_urls[0])
print(image_urls[0])
img = Image.open(BytesIO(img_response.content))
plt.imshow(img)

from PIL import Image import requests from io import BytesIO import matplotlib.pyplot as plt img_response = requests.get(image_urls[0]) print(image_urls[0]) img = Image.open(BytesIO(img_response.content)) plt.imshow(img)

https://storage.googleapis.com/generativeai-downloads/data/scene.jpg

Out[ ]:

<matplotlib.image.AxesImage at 0x128032e40>

在提示词中使用图片进行聊天¶

In [ ]:

response = gemini_pro.chat(messages=[msg])

response = gemini_pro.chat(messages=[msg])

In [ ]:

print(response.message.content)

print(response.message.content)

That's New York City.  More specifically, the photo shows a street in the **SoHo** neighborhood.  The distinctive cast-iron architecture and the pedestrian bridge are characteristic of that area.

支持图片的流式聊天¶

In [ ]:

stream_response = gemini_pro.stream_chat(messages=[msg])

stream_response = gemini_pro.stream_chat(messages=[msg])

In [ ]:

import time


for r in stream_response:
    print(r.delta, end="")
    # Add an artificial wait to make streaming visible in the notebook
    time.sleep(0.5)

import time for r in stream_response: print(r.delta, end="") # Add an artificial wait to make streaming visible in the notebook time.sleep(0.5)

That's New York City.  More specifically, the photo was taken in the **West Village** neighborhood of Manhattan.  The distinctive architecture and the pedestrian bridge are strong clues.

异步支持¶

In [ ]:

response_achat = await gemini_pro.achat(messages=[msg])

response_achat = await gemini_pro.achat(messages=[msg])

In [ ]:

print(response_achat.message.content)

print(response_achat.message.content)

That's New York City.  More specifically, the photo was taken in the **West Village** neighborhood of Manhattan.  The distinctive architecture and the pedestrian bridge are strong clues.

让我们看看如何实现异步流式传输：

In [ ]:

import asyncio

streaming_handler = await gemini_pro.astream_chat(messages=[msg])
async for chunk in streaming_handler:
    print(chunk.delta, end="")
    # Add an artificial wait to make streaming visible in the notebook
    await asyncio.sleep(0.5)

import asyncio streaming_handler = await gemini_pro.astream_chat(messages=[msg]) async for chunk in streaming_handler: print(chunk.delta, end="") # Add an artificial wait to make streaming visible in the notebook await asyncio.sleep(0.5)

That's New York City.  More specifically, the photo was taken in the **West Village** neighborhood of Manhattan.  The distinctive architecture and the pedestrian bridge are strong clues.

完整包含两张图片¶

In [ ]:

image_urls = [
    "https://picsum.photos/id/1/200/300",
    "https://picsum.photos/id/26/200/300",
]


msg = ChatMessage("Is there any relationship between these images?")
for img_url in image_urls:
    msg.blocks.append(ImageBlock(url=img_url))

response_multi = gemini_pro.chat(messages=[msg])

image_urls = [ "https://picsum.photos/id/1/200/300", "https://picsum.photos/id/26/200/300", ] msg = ChatMessage("Is there any relationship between these images?") for img_url in image_urls: msg.blocks.append(ImageBlock(url=img_url)) response_multi = gemini_pro.chat(messages=[msg])

In [ ]:

print(response_multi.message.content)

print(response_multi.message.content)

Yes, there is a relationship between the two images.  Both images depict aspects of a **professional or business-casual lifestyle**.

* **Image 1:** Shows someone working on a laptop, suggesting remote work, freelancing, or a business-related task.

* **Image 2:** Shows a flat lay of accessories commonly associated with a professional or stylish individual: sunglasses, a bow tie, a pen, a watch, glasses, and a phone.  These items suggest a certain level of personal style and preparedness often associated with business or professional settings.

The connection is indirect but thematic.  They both visually represent elements of a similar lifestyle or persona.

第二部分：使用 Gemini + Pydantic 实现图像结构化输出解析¶

• 利用 Gemini 进行图像推理
• 通过 Pydantic 程序从 Gemini 的图像推理结果生成结构化输出

下载示例图像供 Gemini 理解¶

In [ ]:

from pathlib import Path

input_image_path = Path("google_restaurants")
if not input_image_path.exists():
    Path.mkdir(input_image_path)

from pathlib import Path input_image_path = Path("google_restaurants") if not input_image_path.exists(): Path.mkdir(input_image_path)

In [ ]:

!curl -sL "https://docs.google.com/uc?export=download&id=1Pg04p6ss0FlBgz00noHAOAJ1EYXiosKg" -o ./google_restaurants/miami.png
!curl -sL "https://docs.google.com/uc?export=download&id=1dYZy17bD6pSsEyACXx9fRMNx93ok-kTJ" -o ./google_restaurants/orlando.png
!curl -sL "https://docs.google.com/uc?export=download&id=1ShPnYVc1iL_TA1t7ErCFEAHT74-qvMrn" -o ./google_restaurants/sf.png
!curl -sL "https://docs.google.com/uc?export=download&id=1WjISWnatHjwL4z5VD_9o09ORWhRJuYqm" -o ./google_restaurants/toronto.png

!curl -sL "https://docs.google.com/uc?export=download&id=1Pg04p6ss0FlBgz00noHAOAJ1EYXiosKg" -o ./google_restaurants/miami.png !curl -sL "https://docs.google.com/uc?export=download&id=1dYZy17bD6pSsEyACXx9fRMNx93ok-kTJ" -o ./google_restaurants/orlando.png !curl -sL "https://docs.google.com/uc?export=download&id=1ShPnYVc1iL_TA1t7ErCFEAHT74-qvMrn" -o ./google_restaurants/sf.png !curl -sL "https://docs.google.com/uc?export=download&id=1WjISWnatHjwL4z5VD_9o09ORWhRJuYqm" -o ./google_restaurants/toronto.png

定义结构化解析器的 Pydantic 类¶

In [ ]:

from pydantic import BaseModel
from PIL import Image
import matplotlib.pyplot as plt


class GoogleRestaurant(BaseModel):
    """Data model for a Google Restaurant."""

    restaurant: str
    food: str
    location: str
    category: str
    hours: str
    price: str
    rating: float
    review: str
    description: str
    nearby_tourist_places: str


google_image_url = "./google_restaurants/miami.png"
image = Image.open(google_image_url).convert("RGB")

plt.figure(figsize=(16, 5))
plt.imshow(image)

from pydantic import BaseModel from PIL import Image import matplotlib.pyplot as plt class GoogleRestaurant(BaseModel): """Data model for a Google Restaurant.""" restaurant: str food: str location: str category: str hours: str price: str rating: float review: str description: str nearby_tourist_places: str google_image_url = "./google_restaurants/miami.png" image = Image.open(google_image_url).convert("RGB") plt.figure(figsize=(16, 5)) plt.imshow(image)

Out[ ]:

<matplotlib.image.AxesImage at 0x10953cce0>

调用 Pydantic 程序并生成结构化输出¶

In [ ]:

from llama_index.multi_modal_llms.gemini import GeminiMultiModal
from llama_index.core.program import MultiModalLLMCompletionProgram
from llama_index.core.output_parsers import PydanticOutputParser

prompt_template_str = """\
    can you summarize what is in the image\
    and return the answer with json format \
"""


def pydantic_gemini(
    model_name, output_class, image_documents, prompt_template_str
):
    gemini_llm = GeminiMultiModal(model_name=model_name)

    llm_program = MultiModalLLMCompletionProgram.from_defaults(
        output_parser=PydanticOutputParser(output_class),
        image_documents=image_documents,
        prompt_template_str=prompt_template_str,
        multi_modal_llm=gemini_llm,
        verbose=True,
    )

    response = llm_program()
    return response

from llama_index.multi_modal_llms.gemini import GeminiMultiModal from llama_index.core.program import MultiModalLLMCompletionProgram from llama_index.core.output_parsers import PydanticOutputParser prompt_template_str = """\ can you summarize what is in the image\ and return the answer with json format \ """ def pydantic_gemini( model_name, output_class, image_documents, prompt_template_str ): gemini_llm = GeminiMultiModal(model_name=model_name) llm_program = MultiModalLLMCompletionProgram.from_defaults( output_parser=PydanticOutputParser(output_class), image_documents=image_documents, prompt_template_str=prompt_template_str, multi_modal_llm=gemini_llm, verbose=True, ) response = llm_program() return response

通过 Gemini Vision 模型生成 Pydantic 结构化输出¶

In [ ]:

from llama_index.core import SimpleDirectoryReader

google_image_documents = SimpleDirectoryReader(
    "./google_restaurants"
).load_data()

results = []
for img_doc in google_image_documents:
    pydantic_response = pydantic_gemini(
        "models/gemini-1.5-flash",
        GoogleRestaurant,
        [img_doc],
        prompt_template_str,
    )
    # only output the results for miami for example along with image
    if "miami" in img_doc.image_path:
        for r in pydantic_response:
            print(r)
    results.append(pydantic_response)

from llama_index.core import SimpleDirectoryReader google_image_documents = SimpleDirectoryReader( "./google_restaurants" ).load_data() results = [] for img_doc in google_image_documents: pydantic_response = pydantic_gemini( "models/gemini-1.5-flash", GoogleRestaurant, [img_doc], prompt_template_str, ) # only output the results for miami for example along with image if "miami" in img_doc.image_path: for r in pydantic_response: print(r) results.append(pydantic_response)

> Raw output: ```json
{
  "restaurant": "La Mar by Gaston Acurio",
  "food": "Peruvian & fusion",
  "location": "500 Brickell Key Dr, Miami, FL 33131",
  "category": "South American restaurant",
  "hours": "Opens 6PM, Closes 11 PM",
  "price": "$$$",
  "rating": 4.4,
  "review": "Chic waterfront offering Peruvian & fusion fare, plus bars for cocktails, ceviche & anticuchos.",
  "description": "Chic waterfront offering Peruvian & fusion fare, plus bars for cocktails, ceviche & anticuchos.",
  "nearby_tourist_places": "Brickell Key area with scenic views"
}
```

('restaurant', 'La Mar by Gaston Acurio')
('food', 'Peruvian & fusion')
('location', '500 Brickell Key Dr, Miami, FL 33131')
('category', 'South American restaurant')
('hours', 'Opens 6PM, Closes 11 PM')
('price', '$$$')
('rating', 4.4)
('review', 'Chic waterfront offering Peruvian & fusion fare, plus bars for cocktails, ceviche & anticuchos.')
('description', 'Chic waterfront offering Peruvian & fusion fare, plus bars for cocktails, ceviche & anticuchos.')
('nearby_tourist_places', 'Brickell Key area with scenic views')
> Raw output: ```json
{
  "restaurant": "Mythos Restaurant",
  "food": "American fare in a mythic underwater themed spot",
  "location": "6000 Universal Blvd, Orlando, FL 32819, United States",
  "category": "Restaurant",
  "hours": "Open: Closes in 7 hrs, Islands of Adventure",
  "price": "$$",
  "rating": 4.3,
  "review": "Overlooking Universal Studios/Island sea, this mythic underwater themed spot serves American fare.",
  "description": "Dine-in, Delivery",
  "nearby_tourist_places": "Universal Islands, Jurassic Park River Adventure"
}
```

> Raw output: ```json
{
  "restaurant": "Sam's Grill & Seafood Restaurant",
  "food": "Seafood",
  "location": "374 Bush St, San Francisco, CA 94104, United States",
  "category": "Seafood Restaurant",
  "hours": "Open ⋅ Closes 8:30 PM",
  "price": "$$$",
  "rating": 4.4,
  "review": "Modern spin-off adjacent Sam's Grill, for seafood, drinks & happy hour loungey digs with a patio.",
  "description": "Modern spin-off adjacent Sam's Grill, for seafood, drinks & happy hour loungey digs with a patio.",
  "nearby_tourist_places": "Chinatown, San Francisco"
}
```
> Raw output: ```json
{
  "restaurant": "Lobster Port",
  "food": "Seafood restaurant offering lobster, dim sum & Asian fusion dishes",
  "location": "8432 Leslie St, Thornhill, ON L3T 7M6",
  "category": "Seafood",
  "hours": "Open 10pm",
  "price": "$$",
  "rating": 4.0,
  "review": "Elegant, lively venue with a banquet-hall setup",
  "description": "Elegant, lively venue with a banquet-hall setup offering lobster, dim sum & Asian fusion dishes.",
  "nearby_tourist_places": "Nearby tourist places are not explicitly listed in the image but the map shows various points of interest in the surrounding area."
}
```

观察结果:

• Gemini 完美生成了我们为 Pydantic 类所需的所有元信息
• 它还能从 Google Maps 识别出附近的公园

第三部分：构建用于餐厅推荐的多模态 RAG 系统¶

我们的技术栈由 Gemini + LlamaIndex + Pydantic 结构化输出能力组成

构建文本节点以创建向量存储。存储每家餐厅的元数据和描述信息。¶

In [ ]:

from llama_index.core.schema import TextNode

nodes = []
for res in results:
    text_node = TextNode()
    metadata = {}
    for r in res:
        # set description as text of TextNode
        if r[0] == "description":
            text_node.text = r[1]
        else:
            metadata[r[0]] = r[1]
    text_node.metadata = metadata
    nodes.append(text_node)

from llama_index.core.schema import TextNode nodes = [] for res in results: text_node = TextNode() metadata = {} for r in res: # set description as text of TextNode if r[0] == "description": text_node.text = r[1] else: metadata[r[0]] = r[1] text_node.metadata = metadata nodes.append(text_node)

使用 Gemini Embedding 构建密集检索向量库：将餐厅索引为节点存入向量库¶

In [ ]:

from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.embeddings.gemini import GeminiEmbedding
from llama_index.llms.gemini import Gemini
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.core import Settings
from llama_index.core import StorageContext
import qdrant_client


# Create a local Qdrant vector store
client = qdrant_client.QdrantClient(path="qdrant_gemini_3")

vector_store = QdrantVectorStore(client=client, collection_name="collection")

# Using the embedding model to Gemini
Settings.embed_model = GeminiEmbedding(
    model_name="models/embedding-001", api_key=GOOGLE_API_KEY
)
Settings.llm = Gemini(api_key=GOOGLE_API_KEY)

storage_context = StorageContext.from_defaults(vector_store=vector_store)

index = VectorStoreIndex(
    nodes=nodes,
    storage_context=storage_context,
)

from llama_index.core import VectorStoreIndex, StorageContext from llama_index.embeddings.gemini import GeminiEmbedding from llama_index.llms.gemini import Gemini from llama_index.vector_stores.qdrant import QdrantVectorStore from llama_index.core import Settings from llama_index.core import StorageContext import qdrant_client # Create a local Qdrant vector store client = qdrant_client.QdrantClient(path="qdrant_gemini_3") vector_store = QdrantVectorStore(client=client, collection_name="collection") # Using the embedding model to Gemini Settings.embed_model = GeminiEmbedding( model_name="models/embedding-001", api_key=GOOGLE_API_KEY ) Settings.llm = Gemini(api_key=GOOGLE_API_KEY) storage_context = StorageContext.from_defaults(vector_store=vector_store) index = VectorStoreIndex( nodes=nodes, storage_context=storage_context, )

使用 Gemini 合成结果并向用户推荐餐厅¶

In [ ]:

query_engine = index.as_query_engine(
    similarity_top_k=1,
)

response = query_engine.query(
    "recommend a Orlando restaurant for me and its nearby tourist places"
)
print(response)

query_engine = index.as_query_engine( similarity_top_k=1, ) response = query_engine.query( "recommend a Orlando restaurant for me and its nearby tourist places" ) print(response)

For a delightful dining experience, I recommend Mythos Restaurant, known for its American cuisine and unique underwater theme. Overlooking Universal Studios' Inland Sea, this restaurant offers a captivating ambiance. After your meal, explore the nearby tourist attractions such as Universal's Islands of Adventure, Skull Island: Reign of Kong, The Wizarding World of Harry Potter, Jurassic Park River Adventure, and Hollywood Rip Ride Rockit, all located near Mythos Restaurant.