I am trying to make a NLP model that gives the similarity score between 2 JD's and eventually matches it with the database . I tried gpt-2 embeddings, bert pretrained embeddings and more use cosine similarity but the similarity score is not at all good .  Please suggest a way to get better accuracy
Have you tried fine-tuning these embeddings on your data or freezing them and training the final output layers? Or are you using these models out of the box?