使用 OpenVINO 实现本地嵌入¶

OpenVINO™ 是一个用于优化和部署 AI 推理的开源工具包。OpenVINO™ 运行时支持多种硬件设备，包括 x86 和 ARM 架构的 CPU，以及英特尔 GPU。该工具包能够提升计算机视觉、自动语音识别、自然语言处理及其他常见任务中的深度学习性能。

Hugging Face 的嵌入模型可通过 OpenVINOEmbedding 或 OpenVINOGENAIEmbedding 类获得 OpenVINO 支持，而 OpenClip 模型则可通过 OpenVINOClipEmbedding 类实现支持。

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

In [ ]:

%pip install llama-index-embeddings-openvino

%pip install llama-index-embeddings-openvino

In [ ]:

!pip install llama-index

!pip install llama-index

模型导出器¶

您可以使用 create_and_save_openvino_model 函数将模型导出为 OpenVINO IR 格式，并从本地文件夹加载该模型。

In [ ]:

from llama_index.embeddings.huggingface_openvino import OpenVINOEmbedding

OpenVINOEmbedding.create_and_save_openvino_model(
    "BAAI/bge-small-en-v1.5", "./bge_ov"
)

from llama_index.embeddings.huggingface_openvino import OpenVINOEmbedding OpenVINOEmbedding.create_and_save_openvino_model( "BAAI/bge-small-en-v1.5", "./bge_ov" )

/home2/ethan/intel/llama_index/llama_test/lib/python3.10/site-packages/openvino/runtime/__init__.py:10: DeprecationWarning: The `openvino.runtime` module is deprecated and will be removed in the 2026.0 release. Please replace `openvino.runtime` with `openvino`.
  warnings.warn(

Saved OpenVINO model to ./bge_ov. Use it with `embed_model = OpenVINOEmbedding(model_id_or_path='./bge_ov')`.

模型加载¶

若设备配备英特尔 GPU，可通过指定 device="gpu" 参数在其上运行推理。

In [ ]:

ov_embed_model = OpenVINOEmbedding(model_id_or_path="./bge_ov", device="cpu")

ov_embed_model = OpenVINOEmbedding(model_id_or_path="./bge_ov", device="cpu")

In [ ]:

embeddings = ov_embed_model.get_text_embedding("Hello World!")
print(len(embeddings))
print(embeddings[:5])

embeddings = ov_embed_model.get_text_embedding("Hello World!") print(len(embeddings)) print(embeddings[:5])

384
[-0.0030246784444898367, -0.012189766392111778, 0.04163273051381111, -0.037758368998765945, 0.02439723163843155]

使用 OpenVINO GenAI 加载模型¶

为避免运行时对 PyTorch 的依赖，您可以使用 OpenVINOGENAIEmbedding 类加载本地嵌入模型。

In [ ]:

%pip install llama-index-embeddings-openvino-genai

%pip install llama-index-embeddings-openvino-genai

In [ ]:

from llama_index.embeddings.openvino_genai import OpenVINOGENAIEmbedding

ov_embed_model = OpenVINOGENAIEmbedding(model_path="./bge_ov", device="CPU")

from llama_index.embeddings.openvino_genai import OpenVINOGENAIEmbedding ov_embed_model = OpenVINOGENAIEmbedding(model_path="./bge_ov", device="CPU")

/home2/ethan/intel/llama_index/llama_test/lib/python3.10/site-packages/openvino/runtime/__init__.py:10: DeprecationWarning: The `openvino.runtime` module is deprecated and will be removed in the 2026.0 release. Please replace `openvino.runtime` with `openvino`.
  warnings.warn(

In [ ]:

embeddings = ov_embed_model.get_text_embedding("Hello World!")
print(len(embeddings))
print(embeddings[:5])

embeddings = ov_embed_model.get_text_embedding("Hello World!") print(len(embeddings)) print(embeddings[:5])

384
[-0.0030246784444898367, -0.012189766392111778, 0.04163273051381111, -0.037758368998765945, 0.02439723163843155]

OpenClip 模型导出器¶

类 OpenVINOClipEmbedding 支持通过 OpenVINO 运行时导出和加载 open_clip 模型。

In [ ]:

%pip install open_clip_torch

%pip install open_clip_torch

In [ ]:

from llama_index.embeddings.huggingface_openvino import (
    OpenVINOClipEmbedding,
)

OpenVINOClipEmbedding.create_and_save_openvino_model(
    "laion/CLIP-ViT-B-32-laion2B-s34B-b79K",
    "ViT-B-32-ov",
)

from llama_index.embeddings.huggingface_openvino import ( OpenVINOClipEmbedding, ) OpenVINOClipEmbedding.create_and_save_openvino_model( "laion/CLIP-ViT-B-32-laion2B-s34B-b79K", "ViT-B-32-ov", )

多模态模型加载¶

若您的设备配备英特尔GPU，可通过指定device="GPU"参数在其上运行推理。

In [ ]:

ov_clip_model = OpenVINOClipEmbedding(
    model_id_or_path="./ViT-B-32-ov", device="CPU"
)

ov_clip_model = OpenVINOClipEmbedding( model_id_or_path="./ViT-B-32-ov", device="CPU" )

使用 OpenVINO 嵌入图像与查询¶

In [ ]:

from PIL import Image
import requests
from numpy import dot
from numpy.linalg import norm

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)

im.save("logo.jpg")
image_embeddings = ov_clip_model.get_image_embedding("logo.jpg")
print("Image dim:", len(image_embeddings))
print("Image embed:", image_embeddings[:5])

text_embeddings = ov_clip_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 PIL import Image import requests from numpy import dot from numpy.linalg import norm 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) im.save("logo.jpg") image_embeddings = ov_clip_model.get_image_embedding("logo.jpg") print("Image dim:", len(image_embeddings)) print("Image embed:", image_embeddings[:5]) text_embeddings = ov_clip_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)

Image:

Image dim: 512
Image embed: [-0.03019799292087555, -0.09727513045072556, -0.6659489274024963, -0.025658488273620605, 0.05379948765039444]
Text dim: 512
Text embed: [-0.15816599130630493, -0.25564345717430115, 0.22376027703285217, -0.34983670711517334, 0.31968361139297485]
Cosine similarity: 0.27307014923203976

如需了解更多信息，请参考：

• OpenVINO 大语言模型指南

• OpenVINO 文档中心

• OpenVINO 入门指南

• LlamaIndex 的 RAG 应用示例