With just two main components you can create a fully working semantic search app. The client performs the heavy work of embedding calculation and simply sends the vector to Qdrant (vector database with built-in API) performing the actual search. Note that the model I'm using in the frontend is quantized (weighs only 30mb) and hence outputs slightly different results in comparison to the original model.
Show HN: Semantic Search with Qdrant and Transformers.js | Heykuki News