LlamaIndex 支持 HuggingFace 嵌入模型，包括 BGE、Mixedbread、Nomic、Jina、E5 等 Sentence Transformer 模型。我们可以使用这些模型为文档和查询生成嵌入，用于检索。

此外，我们还提供了一些工具，可使用 HuggingFace 的 Optimum 库 创建和使用 ONNX 与 OpenVINO 模型。

HuggingFaceEmbedding

基类 HuggingFaceEmbedding 是任意 HuggingFace 嵌入模型的通用封装。Hugging Face 上的所有嵌入模型 都应该可用。你可以参考嵌入模型排行榜 获取更多推荐。

该类依赖 sentence-transformers 包，可通过 pip install sentence-transformers 安装。

注意：如果你之前使用的是 LangChain 中的 HuggingFaceEmbeddings，这里应能得到等效的结果。

如果你在 Colab 中打开这个 Notebook，可能需要先安装 LlamaIndex 🦙。

%pip install llama-index-embeddings-huggingface

!pip install llama-index

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# loads https://huggingface.co/BAAI/bge-small-en-v1.5
embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")

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

384
[-0.003275700844824314, -0.011690810322761536, 0.041559211909770966, -0.03814814239740372, 0.024183044210076332]

基准测试

我们用一个经典的大型文档来比较——IPCC 气候报告第 3 章。

!curl https://www.ipcc.ch/report/ar6/wg2/downloads/report/IPCC_AR6_WGII_Chapter03.pdf --output IPCC_AR6_WGII_Chapter03.pdf

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed

  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
100 20.7M  100 20.7M    0     0  69.6M      0 --:--:-- --:--:-- --:--:-- 70.0M

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.core import Settings

documents = SimpleDirectoryReader(
    input_files=["IPCC_AR6_WGII_Chapter03.pdf"]
).load_data()

基础 HuggingFace 嵌入模型

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# loads BAAI/bge-small-en-v1.5 with the default torch backend
embed_model = HuggingFaceEmbedding(
    model_name="BAAI/bge-small-en-v1.5",
    device="cpu",
    embed_batch_size=8,
)
test_embeds = embed_model.get_text_embedding("Hello World!")

Settings.embed_model = embed_model

%%timeit -r 1 -n 1
index = VectorStoreIndex.from_documents(documents, show_progress=True)

Parsing nodes: 100%|██████████| 172/172 [00:00<00:00, 428.44it/s]
Generating embeddings: 100%|██████████| 459/459 [00:19<00:00, 23.32it/s]

20.2 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

ONNX 嵌入模型

# pip install sentence-transformers[onnx]

# loads BAAI/bge-small-en-v1.5 with the onnx backend
embed_model = HuggingFaceEmbedding(
    model_name="BAAI/bge-small-en-v1.5",
    device="cpu",
    backend="onnx",
    model_kwargs={
        "provider": "CPUExecutionProvider"
    },  # For ONNX, you can specify the provider, see https://sbert.net/docs/sentence_transformer/usage/efficiency.html
)
test_embeds = embed_model.get_text_embedding("Hello World!")

Settings.embed_model = embed_model

%%timeit -r 1 -n 1
index = VectorStoreIndex.from_documents(documents, show_progress=True)

Parsing nodes: 100%|██████████| 172/172 [00:00<00:00, 421.63it/s]
Generating embeddings: 100%|██████████| 459/459 [00:31<00:00, 14.53it/s]

32.1 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

OpenVINO 嵌入模型

# pip install sentence-transformers[openvino]

# loads BAAI/bge-small-en-v1.5 with the openvino backend
embed_model = HuggingFaceEmbedding(
    model_name="BAAI/bge-small-en-v1.5",
    device="cpu",
    backend="openvino",  # OpenVINO is very strong on CPUs
    revision="refs/pr/16",  # BAAI/bge-small-en-v1.5 itself doesn't have an OpenVINO model currently, but there's a PR with it that we can load: https://huggingface.co/BAAI/bge-small-en-v1.5/discussions/16
    model_kwargs={
        "file_name": "openvino_model_qint8_quantized.xml"
    },  # If we're using an optimized/quantized model, we need to specify the file name like this
)
test_embeds = embed_model.get_text_embedding("Hello World!")

Settings.embed_model = embed_model

%%timeit -r 1 -n 1
index = VectorStoreIndex.from_documents(documents, show_progress=True)

Parsing nodes: 100%|██████████| 172/172 [00:00<00:00, 403.15it/s]
Generating embeddings: 100%|██████████| 459/459 [00:08<00:00, 53.83it/s]

9.03 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

参考

• 本地嵌入模型 进一步介绍了如何使用这类本地模型。
• Sentence Transformers > 加速推理 包含大量关于如何有效使用后端选项的文档，包括针对 ONNX 和 OpenVINO 的优化与量化。