使用 HuggingFace 的本地嵌入¶

LlamaIndex 支持 HuggingFace 嵌入模型，包括 BGE、Instructor 等。

此外，我们提供实用工具，可使用 HuggingFace 的 Optimum 库创建和使用 ONNX 模型。

HuggingFaceEmbedding¶

基础的 HuggingFaceEmbedding 类是对任何 HuggingFace 嵌入模型进行的通用包装。Hugging Face 上的所有嵌入模型都应该可以使用。您可以参考嵌入排行榜以获取更多推荐。

此类依赖于 sentence-transformers 包，您可以使用 pip install sentence-transformers 进行安装。

注意：如果您之前使用的是 LangChain 中的 HuggingFaceEmbeddings，这将给出等效的结果。

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

输入 [ ]

%pip install llama-index-embeddings-huggingface
%pip install llama-index-embeddings-instructor

%pip install llama-index-embeddings-huggingface %pip install llama-index-embeddings-instructor

输入 [ ]

!pip install llama-index

!pip install llama-index

输入 [ ]

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# loads BAAI/bge-small-en
# embed_model = HuggingFaceEmbedding()

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

from llama_index.embeddings.huggingface import HuggingFaceEmbedding # 加载 BAAI/bge-small-en # embed_model = HuggingFaceEmbedding() # 加载 BAAI/bge-small-en-v1.5 embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")

/home/loganm/miniconda3/envs/llama-index/lib/python3.11/site-packages/torch/cuda/__init__.py:546: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")

输入 [ ]

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

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

Hello World!
384
[-0.030880315229296684, -0.11021008342504501, 0.3917851448059082, -0.35962796211242676, 0.22797748446464539]

InstructorEmbedding¶

Instructor Embeddings 是一类专门训练用于根据指令增强其嵌入的模型。默认情况下，查询会被赋予 query_instruction="Represent the question for retrieving supporting documents: "，文本会被赋予 text_instruction="Represent the document for retrieval: "。

它们依赖于 Instructor 和 SentenceTransformers (版本 2.2.2) pip 包，您可以使用 pip install InstructorEmbedding 和 pip install -U sentence-transformers==2.2.2 进行安装。

输入 [ ]

from llama_index.embeddings.instructor import InstructorEmbedding

embed_model = InstructorEmbedding(model_name="hkunlp/instructor-base")

from llama_index.embeddings.instructor import InstructorEmbedding embed_model = InstructorEmbedding(model_name="hkunlp/instructor-base")

/home/loganm/miniconda3/envs/llama-index/lib/python3.11/site-packages/InstructorEmbedding/instructor.py:7: TqdmExperimentalWarning: Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)
  from tqdm.autonotebook import trange

load INSTRUCTOR_Transformer

/home/loganm/miniconda3/envs/llama-index/lib/python3.11/site-packages/torch/cuda/__init__.py:546: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")

max_seq_length  512

输入 [ ]

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

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

768
[ 0.02155361 -0.06098218  0.01796207  0.05490903  0.01526906]

OptimumEmbedding¶

Optimum 是 HuggingFace 的一个库，用于以 ONNX 格式导出和运行 HuggingFace 模型。

您可以使用 pip install transformers optimum[exporters] 安装依赖项。

首先，我们需要创建 ONNX 模型。ONNX 模型提供了更高的推理速度，并且可以跨平台使用（例如在 TransformersJS 中）。

输入 [ ]

from llama_index.embeddings.huggingface_optimum import OptimumEmbedding

OptimumEmbedding.create_and_save_optimum_model(
    "BAAI/bge-small-en-v1.5", "./bge_onnx"
)

from llama_index.embeddings.huggingface_optimum import OptimumEmbedding OptimumEmbedding.create_and_save_optimum_model( "BAAI/bge-small-en-v1.5", "./bge_onnx" )

/home/loganm/miniconda3/envs/llama-index/lib/python3.11/site-packages/torch/cuda/__init__.py:546: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
Framework not specified. Using pt to export to ONNX.
Using the export variant default. Available variants are:
	- default: The default ONNX variant.
Using framework PyTorch: 2.0.1+cu117
Overriding 1 configuration item(s)
	- use_cache -> False

============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Saved optimum model to ./bge_onnx. Use it with `embed_model = OptimumEmbedding(folder_name='./bge_onnx')`.

输入 [ ]

embed_model = OptimumEmbedding(folder_name="./bge_onnx")

embed_model = OptimumEmbedding(folder_name="./bge_onnx")

输入 [ ]

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

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

384
[-0.10364960134029388, -0.20998482406139374, -0.01883639395236969, -0.5241696834564209, 0.0335749015212059]

基准测试 (Benchmarking)¶

让我们尝试比较使用一个经典的文档——IPCC 气候报告，第三章。

输入 [ ]

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

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

huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 20.7M  100 20.7M    0     0  16.5M      0  0:00:01  0:00:01 --:--:-- 16.5M

输入 [ ]

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

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

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¶

输入 [ ]

import os
import openai

# needed to synthesize responses later
os.environ["OPENAI_API_KEY"] = "sk-..."
openai.api_key = os.environ["OPENAI_API_KEY"]

import os import openai # 用于后续合成响应 os.environ["OPENAI_API_KEY"] = "sk-..." openai.api_key = os.environ["OPENAI_API_KEY"]

输入 [ ]

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# loads BAAI/bge-small-en-v1.5
embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")
test_emeds = embed_model.get_text_embedding("Hello World!")

Settings.embed_model = embed_model

from llama_index.embeddings.huggingface import HuggingFaceEmbedding # 加载 BAAI/bge-small-en-v1.5 embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5") test_emeds = embed_model.get_text_embedding("Hello World!") Settings.embed_model = embed_model

输入 [ ]

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

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

Parsing documents into nodes:   0%|          | 0/172 [00:00<?, ?it/s]

Generating embeddings:   0%|          | 0/428 [00:00<?, ?it/s]

1min 27s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

Optimum 嵌入¶

我们可以使用之前创建的 onnx 嵌入

输入 [ ]

from llama_index.embeddings.huggingface_optimum import OptimumEmbedding

embed_model = OptimumEmbedding(folder_name="./bge_onnx")
test_emeds = embed_model.get_text_embedding("Hello World!")

Settings.embed_model = embed_model

from llama_index.embeddings.huggingface_optimum import OptimumEmbedding embed_model = OptimumEmbedding(folder_name="./bge_onnx") test_emeds = embed_model.get_text_embedding("Hello World!") Settings.embed_model = embed_model

输入 [ ]

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

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

Parsing documents into nodes:   0%|          | 0/172 [00:00<?, ?it/s]

Generating embeddings:   0%|          | 0/428 [00:00<?, ?it/s]

1min 9s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)