There's no easy way to index ColBert multi-vectors in a scalable way that I know of. Vespa seems to rely heavily on binary quantization, which can cost a lot in recall loss. And for most cases, using ColBert as a reranker is good enough, as the pgvector example you posted.