Bi-Encoder for 1C:Enterprise Code Retrieval Using Synthetic Triplets

A specialized bi-encoder achieves 0.5992 balanced macro nDCG@10 for 1C:Enterprise code retrieval by fine-tuning on 784,057 synthetic triplets generated via Gemma-4-26B. The method utilizes Matryoshka Representation Learning and a privacy-aware tokenizer to handle domain-specific Russian syntax and terminology.