I strongly advocate this! If you're starting off in this space, check if this barebones implementation isn't all you need. You can't beat the accuracy of a full search; in theory, you're trading off scalability, but validate if you need the scale where this tradeoff begins to show. And yes, sbert is great, and it gives you options to choose [1] between accuracy (MPNet) and speed (MiniLM). There are also multi-lingual options. And remember, you can also fine-tune MPNet with SetFit. And there are always new and interesting embeddings being released, so remember to re-assess the fitment of embeddings once in a while against what you're using, e.g., LLM2vec or ModernBERT. A good idea would be to keep checking MTEB [2].

[1] https://www.sbert.net/docs/sentence_transformer/pretrained_m...

[2] https://huggingface.co/spaces/mteb/leaderboard