SberTech has detailed an AI agent that checks technical documentation against a glossary inside an IDE and proposes edits for the writer to accept or reject. Its initial implementation combined the Continue IDE plugin, the Langflow visual builder and an internal Qwen3-32b model. The team later replaced the low-code workflow with its own Python implementation as it prepared the system for production use.
The agent is designed to standardize terminology rather than produce documentation from scratch. A client agent sends selected text to a server-side component, which checks it against the glossary and returns a revised version. SberTech says the glossary, published in a GitVerse repository, now contains more than 750 terms and is still being expanded by documentation developers and cybersecurity specialists.
The team assembled its first prototype in several hours, but larger inputs exposed practical limits. According to SberTech, the model received oversized glossary sections, sometimes lost track of earlier corrections, responded slowly and consumed excessive GPU resources. The developers also considered Langflow’s visual-block implementation insufficiently reliable for a production environment.
To reduce the material placed in the model’s context, SberTech introduced vector search. Terms, approved alternatives, English forms and discouraged synonyms are embedded with bge-m3 and stored in Qdrant. The input is split into words, each word is matched against stored vectors, and relevant entries are deduplicated and assembled into a text-specific mini-glossary. The experimental similarity threshold is 0.85–0.90, although the team says the appropriate balance depends heavily on the embedding model.
The vector store runs locally in memory through qdrant-client, without a separate Kubernetes cluster. Because Langflow lacked the required logic out of the box, the team rewrote the workflow in Python in one day and then spent another week debugging it. It also converted the source glossary from inconsistently formatted Markdown into YAML to simplify machine parsing.
Practical context: Selecting only relevant glossary entries makes the workflow better suited to interactive checking than sending the entire reference set with every request. It does not remove the need for review, however: SberTech reports that model responses can introduce unnecessary information and that formatting can become unstable after edits.
SberTech did not announce a public release of the agent. The next planned stage is to check text against the company’s internal editorial policy as well as its glossary. The team is also developing a full MCP server with an API and intends to package the logic as separate skills that could run as a script, a local MCP server or through an API.
Sources
Event date: 2026-09-23. Primary source date: 2026-09-23.
