How To: Vector embeddings and semantic similarity WITHOUT OpenAI

Why avoid OpenAI

1. OpenAI is great, but costs money.

2. Learning to do things with open source models is a valuable skill. Open source models are the future, Google recently said "We Have No Moat, And Neither Does OpenAI"

SentenceTransformerService

I'm just going to share the code.

If you have any questions, plug them into ChatGPT and maybe share the answers in comments below.

First, install the sentence-transformers and torch libraries.

pip install sentence-transformers 
pip install torch

Then, you can use them like so.

from sentence_transformers import SentenceTransformer, util # type: ignore
from torch import Tensor

class SentenceTransformerService:
    """
    A service that uses the MiniLM model to convert text to embeddings and then compares the 
    embeddings to compute similarity
    """

    def __init__(self):
        # The MiniLM model converts text to 384-dimensional embeddings
        self.model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

    def compare(self, text1: str, text2: str) -> float:
        # Convert the two sentences into two embeddings
        embedding_1: Tensor = self.model.encode(text1, convert_to_tensor=True) # type: ignore
        embedding_2: Tensor = self.model.encode(text2, convert_to_tensor=True) # type: ignore

        # Compute the cosine similarity between the two embeddings
        similarity = util.pytorch_cos_sim(embedding_1, embedding_2)[0][0].item()

        # Returns a float from 0.00 to 1.00, where 1.00 means the two sentences are identical
        return similarity

The results

The output is as follows.

As you can see, "hello" is equal to "hello", so the result is exactly 1.0

"Hello" and "hi" are also pretty similar, so the result is ~0.81

However, "hello" and "world" are very dissimilar, so the result is only ~0.35

sts = SentenceTransformerService()
print(sts.compare("hello", "hello")) # Output: 1.0
print(sts.compare("hello", "hi")) # Output: 0.8071529865264893
print(sts.compare("hello", "hey")) # Output: 0.702325701713562
print(sts.compare("hello", "good day")) # Output: 0.5354946255683899
print(sts.compare("hello", "hola")) # Output: 0.42029112577438354
print(sts.compare("hello", "world")) # Output: 0.34536516666412354

Not bad for a local model, eh?

Best of luck!