Supercharging Vector Search with ColBERT Reranking in PostgreSQL

Traditional vector search relies on sentence embeddings, potentially losing fine-grained details. ColBERT overcomes this by representing text as token-level multi-vectors, retaining nuanced information and improving accuracy. However, token-level interaction is computationally expensive. This blog post demonstrates combining sentence-level vector search with ColBERT token-level reranking using the PostgreSQL extensions VectorChord and pgvector. This approach performs a fast initial search using sentence embeddings, followed by reranking with ColBERT for improved results. Significant improvements were observed on several BEIR datasets.