Fine-tuning Vs Embedding for Question Answering

I am looking to develop a question answering system to work on structured data. I have the option of either using the Embeddings endpoint to store the data as vectors and then use embeddings similar to embeddings of the query to select the top-k results to be returned and use a verification method to get the correct answer, or use fine-tuning to train the model to the data and then make use of the fine-tuned model.

Considering that the data could change with time, can someone please guide me how to decide which alternative would be best to use?

Embeddings work well if the length of the text is not short.

Otherwise, a traditional full-text DB search works better on short strings.