WhereIsAI

/

UAE-Large-V1

like 211

Follow

WhereIsAI 10

Thanks for following our work.

First, our model has a token limit of 512, and you don't need to specify the max token explicitly.

Second, if you use it for retrieval, you need to set a prompt for the query (no need for documents). There are two ways:

• angle.set_prompt(Prompts.C); angle.encode(query)
• manually apply the prompt to the query text when calling the encode() function. For example angle.encode(Prompts.C.format(text=query)).

For other non-retrieval scenarios, there is no need to set the prompt.

If you use it for retrieval and summarization at the same time, I suggest manually setting the prompt when calling encode(), i.e., the second way.

Third, currently, angle_emb does not support onnx inference. You can use HuggingFace optimum for inference.
Here is the usage: https://huggingface.co/WhereIsAI/UAE-Large-V1/discussions/10#659df9c060736ff2a6dc04e7

Thanks.
Regarding onnx, I'm getting some unexpected results.
The onnx is almost 4X slower than the regular embed.

Here's my code. Any idea what might be wrong? I'm running this on a T4 in colab.

!pip install --upgrade optimum[onnxruntime] onnxruntime-gpu onnx

from optimum.onnxruntime import ORTModelForFeatureExtraction
from optimum.pipelines import pipeline

model = ORTModelForFeatureExtraction.from_pretrained('WhereIsAI/UAE-Large-V1', file_name="onnx/model.onnx")

def onnx_embed(query):
  extractor = pipeline('feature-extraction', model=model)
  output = extractor(query)
  return output

import time
query = " some test"
timh = time.time()
avec = onnx_embed(query)[0][0]
print(time.time()-timh) # prints 0.28627943992614746
print(avec)

!pip install -U angle-emb --quiet
from angle_emb import AnglE, Prompts
angle = AnglE.from_pretrained('WhereIsAI/UAE-Large-V1', pooling_strategy='cls').cuda()

def uae_embed(query):
 vec = angle.encode(query, to_numpy=True)
  return vec.flatten().tolist()

import time
query = " some test"
timh = time.time()
vec = uae_embed(query)
print(vec)
print(time.time()-timh) # prints 0.07494688034057617

Also wanted a bit of advice on one particular use case. I'm trying to do extractive summarization in a passage and find the sentence in the passage that is most similar to the whole passage.
So for example, in the passage: The ship set out to sail around the world. It was manned by sailors from around the world. They set out to sail yesterday."
It would choose let's say "The ship set out to sail around the world" as the sentence most similar to the whole passage.

In this case, I am essentially comparing 1 passage to 3 individual sentences. Would this be considered retrieval as well? Do I have to add the query prompt to the passage embedding here, but not add them to the individual sentence embeddings?