Jina 嵌入模型¶

如果您在 Colab 上打开此 Notebook，可能需要安装 LlamaIndex 🦙。

In [ ]:

%pip install llama-index-embeddings-jinaai
%pip install llama-index-llms-openai

%pip install llama-index-embeddings-jinaai %pip install llama-index-llms-openai

In [ ]:

!pip install llama-index

!pip install llama-index

您可能还需要其他未随 llama-index 直接提供的软件包

In [ ]:

!pip install Pillow

!pip install Pillow

在本示例中，您需要一个 API 密钥，可从 https://jina.ai/embeddings/ 获取

In [ ]:

# Initilise with your api key
import os

jinaai_api_key = "YOUR_JINAAI_API_KEY"
os.environ["JINAAI_API_KEY"] = jinaai_api_key

# Initilise with your api key import os jinaai_api_key = "YOUR_JINAAI_API_KEY" os.environ["JINAAI_API_KEY"] = jinaai_api_key

通过 JinaAI API 使用 Jina 嵌入模型处理文本与查询嵌入¶

您可以使用 JinaEmbedding 类对文本和查询进行编码。Jina 提供了一系列适应不同使用场景的模型。

模型	维度	语言	套娃表示学习 (MRL)	上下文长度
jina-embeddings-v3	1024	多语言（89种）	支持	8192
jina-embeddings-v2-base-en	768	英语	不支持	8192
jina-embeddings-v2-base-de	768	德语 & 英语	不支持	8192
jina-embeddings-v2-base-es	768	西班牙语 & 英语	不支持	8192
jina-embeddings-v2-base-zh	768	中文 & 英语	不支持	8192

推荐模型：jina-embeddings-v3

我们推荐使用 jina-embeddings-v3 作为 Jina AI 最新且性能最强的嵌入模型。该模型在基础架构上训练了5个针对特定任务的适配器，可优化各类嵌入使用场景。

默认情况下 JinaEmbedding 类使用 jina-embeddings-v3。该模型在基础架构之外，还针对不同嵌入用途训练了5个任务专用适配器。

任务专用适配器：

在请求中包含 task 参数以优化下游应用：

• retrieval.query：用于检索任务中对用户查询或问题进行编码
• retrieval.passage：用于检索任务索引阶段对长文档进行编码
• classification：用于文本分类任务的文本编码
• text-matching：用于相似度匹配的文本编码（如测量两个句子的相似度）
• separation：适用于聚类或重排序任务

套娃表示学习（MRL）：

jina-embeddings-v3 支持套娃表示学习技术，允许用户在最小性能损失下控制嵌入维度。
在请求中包含 dimensions 参数来选择所需维度。
默认 dimensions 设为1024，推荐值介于256至1024之间。
可参考下表了解维度与性能的关系：

维度	32	64	128	256	512	768	1024
平均检索性能 (nDCG@10)	52.54	58.54	61.64	62.72	63.16	63.3	63.35

长上下文嵌入模型的延迟分块技术

jina-embeddings-v3 支持延迟分块技术，该技术能利用模型的长上下文能力生成上下文分块嵌入。在请求中包含 late_chunking=True 即可启用上下文分块表示。当设为true时，Jina AI API会将输入字段中的所有句子拼接为单个字符串输入模型。模型内部会先对这个长拼接字符串进行嵌入，然后执行延迟分块操作，最终返回与输入列表长度匹配的嵌入列表。

In [ ]:

from llama_index.embeddings.jinaai import JinaEmbedding

text_embed_model = JinaEmbedding(
    api_key=jinaai_api_key,
    model="jina-embeddings-v3",
    # choose `retrieval.passage` to get passage embeddings
    task="retrieval.passage",
)

embeddings = text_embed_model.get_text_embedding("This is the text to embed")
print("Text dim:", len(embeddings))
print("Text embed:", embeddings[:5])

query_embed_model = JinaEmbedding(
    api_key=jinaai_api_key,
    model="jina-embeddings-v3",
    # choose `retrieval.query` to get query embeddings, or choose your desired task type
    task="retrieval.query",
    # `dimensions` allows users to control the embedding dimension with minimal performance loss. by default it is 1024.
    # A number between 256 and 1024 is recommended.
    dimensions=512,
)

embeddings = query_embed_model.get_query_embedding(
    "This is the query to embed"
)
print("Query dim:", len(embeddings))
print("Query embed:", embeddings[:5])

from llama_index.embeddings.jinaai import JinaEmbedding text_embed_model = JinaEmbedding( api_key=jinaai_api_key, model="jina-embeddings-v3", # choose `retrieval.passage` to get passage embeddings task="retrieval.passage", ) embeddings = text_embed_model.get_text_embedding("This is the text to embed") print("Text dim:", len(embeddings)) print("Text embed:", embeddings[:5]) query_embed_model = JinaEmbedding( api_key=jinaai_api_key, model="jina-embeddings-v3", # choose `retrieval.query` to get query embeddings, or choose your desired task type task="retrieval.query", # `dimensions` allows users to control the embedding dimension with minimal performance loss. by default it is 1024. # A number between 256 and 1024 is recommended. dimensions=512, ) embeddings = query_embed_model.get_query_embedding( "This is the query to embed" ) print("Query dim:", len(embeddings)) print("Query embed:", embeddings[:5])

通过 JinaAI API 使用 Jina CLIP 嵌入图像与查询¶

您也可以使用 JinaEmbedding 类对图像和查询进行编码

In [ ]:

from llama_index.embeddings.jinaai import JinaEmbedding
from PIL import Image
import requests
from numpy import dot
from numpy.linalg import norm

embed_model = JinaEmbedding(
    api_key=jinaai_api_key,
    model="jina-clip-v1",
)

image_url = "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcStMP8S3VbNCqOQd7QQQcbvC_FLa1HlftCiJw&s"
im = Image.open(requests.get(image_url, stream=True).raw)
print("Image:")
display(im)

image_embeddings = embed_model.get_image_embedding(image_url)
print("Image dim:", len(image_embeddings))
print("Image embed:", image_embeddings[:5])

text_embeddings = embed_model.get_text_embedding(
    "Logo of a pink blue llama on dark background"
)
print("Text dim:", len(text_embeddings))
print("Text embed:", text_embeddings[:5])

cos_sim = dot(image_embeddings, text_embeddings) / (
    norm(image_embeddings) * norm(text_embeddings)
)
print("Cosine similarity:", cos_sim)

from llama_index.embeddings.jinaai import JinaEmbedding from PIL import Image import requests from numpy import dot from numpy.linalg import norm embed_model = JinaEmbedding( api_key=jinaai_api_key, model="jina-clip-v1", ) image_url = "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcStMP8S3VbNCqOQd7QQQcbvC_FLa1HlftCiJw&s" im = Image.open(requests.get(image_url, stream=True).raw) print("Image:") display(im) image_embeddings = embed_model.get_image_embedding(image_url) print("Image dim:", len(image_embeddings)) print("Image embed:", image_embeddings[:5]) text_embeddings = embed_model.get_text_embedding( "Logo of a pink blue llama on dark background" ) print("Text dim:", len(text_embeddings)) print("Text embed:", text_embeddings[:5]) cos_sim = dot(image_embeddings, text_embeddings) / ( norm(image_embeddings) * norm(text_embeddings) ) print("Cosine similarity:", cos_sim)

批量嵌入¶

您还可以批量嵌入文本，通过设置 embed_batch_size 参数控制批量大小（若不传入则默认值为 10，且不应超过 2048）

In [ ]:

embed_model = JinaEmbedding(
    api_key=jinaai_api_key,
    model="jina-embeddings-v3",
    embed_batch_size=16,
    task="retrieval.passage",
)

embeddings = embed_model.get_text_embedding_batch(
    ["This is the text to embed", "More text can be provided in a batch"]
)

print(len(embeddings))
print(embeddings[0][:5])

embed_model = JinaEmbedding( api_key=jinaai_api_key, model="jina-embeddings-v3", embed_batch_size=16, task="retrieval.passage", ) embeddings = embed_model.get_text_embedding_batch( ["This is the text to embed", "More text can be provided in a batch"] ) print(len(embeddings)) print(embeddings[0][:5])

使用 Jina AI 嵌入构建 RAG 流水线¶

下载数据¶

In [ ]:

!mkdir -p 'data/paul_graham/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/paul_graham/paul_graham_essay.txt' -O 'data/paul_graham/paul_graham_essay.txt'

!mkdir -p 'data/paul_graham/' !wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/paul_graham/paul_graham_essay.txt' -O 'data/paul_graham/paul_graham_essay.txt'

导入¶

In [ ]:

import logging
import sys

logging.basicConfig(stream=sys.stdout, level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader

from llama_index.llms.openai import OpenAI
from llama_index.core.response.notebook_utils import display_source_node

from IPython.display import Markdown, display

import logging import sys logging.basicConfig(stream=sys.stdout, level=logging.INFO) logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout)) from llama_index.core import VectorStoreIndex, SimpleDirectoryReader from llama_index.llms.openai import OpenAI from llama_index.core.response.notebook_utils import display_source_node from IPython.display import Markdown, display

加载数据¶

In [ ]:

documents = SimpleDirectoryReader("./data/paul_graham/").load_data()

documents = SimpleDirectoryReader("./data/paul_graham/").load_data()

构建索引¶

In [ ]:

your_openai_key = "YOUR_OPENAI_KEY"
llm = OpenAI(api_key=your_openai_key)
embed_model = JinaEmbedding(
    api_key=jinaai_api_key,
    model="jina-embeddings-v3",
    embed_batch_size=16,
    task="retrieval.passage",
)

index = VectorStoreIndex.from_documents(
    documents=documents, embed_model=embed_model
)

your_openai_key = "YOUR_OPENAI_KEY" llm = OpenAI(api_key=your_openai_key) embed_model = JinaEmbedding( api_key=jinaai_api_key, model="jina-embeddings-v3", embed_batch_size=16, task="retrieval.passage", ) index = VectorStoreIndex.from_documents( documents=documents, embed_model=embed_model )

构建检索器¶

In [ ]:

search_query_retriever = index.as_retriever()

search_query_retrieved_nodes = search_query_retriever.retrieve(
    "What happened after the thesis?"
)

search_query_retriever = index.as_retriever() search_query_retrieved_nodes = search_query_retriever.retrieve( "What happened after the thesis?" )

In [ ]:

for n in search_query_retrieved_nodes:
    display_source_node(n, source_length=2000)

for n in search_query_retrieved_nodes: display_source_node(n, source_length=2000)