The Bitrix24 DevRel team has developed DocFlow, a service that automates the translation of technical documentation from Russian to English. The system is designed to minimize manual correction by technical writers, transforming the translation process into a managed pipeline that parses files, translates specific segments, and automatically verifies the results.
To prevent AI from breaking the document structure, the team implemented a deterministic layer of rules and checks. The system uses placeholders like [[PROTECTED_0]] to hide sensitive elements such as API methods, code examples, links, and service markup in YFM (Yandex Flavored Markdown) format from the model. These original fragments are stored in memory and restored after the translation is complete.
The service handles complex elements separately: it extracts Russian text from within code (JavaScript, PHP, JSON) for translation while keeping the syntax intact. Tables are processed by translating only the cell content, preserving the structural markers like #|, ||, and |#. Additionally, a translation memory and a glossary of terms (e.g., 'portal' as 'account') are used to ensure consistency and reduce costs.
To maintain reproducibility, the model's temperature is set to 0. The system also includes a validator and a series of 'fixers' to correct common errors. For example, the CellTypeFixer restores data type lines in YFM cells based on the original document. If the system detects a missing placeholder after translation, it sends a repeat request to the model to recover the lost element.
