Integrate semantic/vector search using BERT-based models

Most likely not as a replacement, but as an extension to our current full-text search, we should investigate integrating semantic/vector search based on document embeddings which can be created using BERT-based models available via e.g. https://sbert.net/ or https://github.com/guillaume-be/rust-bert. This can be integrated with the scoring in the existing index or as a separate index like https://qdrant.tech/.