Abstract: In diverse applications like image clustering, facial recognition and text embeddings, similarity search is critical. Deep models utilize feature embeddings for efficient representation, learning shared similarities during training. However, developing new models raises compatibility issues, necessitating the re-extraction of the entire embedding database (backfilling). Very large datasets become a bigger problem, considering the necessary time and computational power. While solutions like backward compatibility models and harmonic embeddings exists, they may be impractical without the original input data. We present a simple yet efficient model called FC-Aligner, which converts embeddings from a previous embedding space to a new one through a regression-inspired approach. We use production data from a real-world face recognition system, with a total of $3.39$M samples. Results show an acceptable increase in FAR from $0.0032\%$ to $0.0048\%$, while keeping a similar FRR (${\sim}5\%$). Using FC-Aligner in $20$M embeddings is $11\times$ faster and $2.5\times$ cheaper than backfilling.