Sber has launched DAIS, an internal AI agent that helps designers and developers search the FinAI design system’s documentation. Built around GigaChat-2-PRO, the service answers questions about components, code and editorial policy, with links to the underlying documents. Employees are testing it; Sber’s account does not announce external availability.
The project began as a way to make an eight-section editorial policy easier to use, before expanding across the design system. Its knowledge base covers more than 50 files on atomic components, 25 files about the table component, over 20 files for other template components, more than 20 design guides and the editorial policy.
Sber says preparing that knowledge base required more effort than writing the agent. The team exported material from Confluence and Pixso in DOCX, PDF and Sketch formats, converted it to Markdown, and checked each resulting file against the design and code. GigaCode reportedly made the Markdown conversion three to four times faster, but human review and a second comparison with the source remained mandatory.
DAIS uses Python 3.12, FastAPI 0.115.5, GigaChat-2-PRO and retrieval-augmented generation with OpenSearch. The documentation was divided into roughly 1,000 small chunks. For each question, the system retrieves relevant passages, takes the top seven results, requests their complete sections and sends the question and assembled context to GigaChat before returning an answer with a source link.
From June 17 through September 8, users submitted 663 requests: 410 about development, 160 about editorial policy and 93 about design. According to the team, almost 60% received a specific answer from the documentation. The success rate for table-component questions was about 44%, partly because developers often asked about undocumented edge cases or advanced properties.
The team says DAIS correctly interprets a user’s description only about half the time, performs better on general questions and can lose context during long sequences of follow-ups. When its knowledge base lacks an answer, it is designed to direct users to Storybook, Pixso guides or the design-system support chat rather than fill the gap with unsupported information.
Practical context: The case suggests that a corporate RAG assistant depends heavily on maintained source material, consistent metadata, reference FAQs and verifiable links. The published figures do not yet establish an economic benefit: Sber has not reported reductions in search time or support workload. Planned work includes handling layouts through MCP, integration with support chats and automated knowledge-base updates.
| Category | Number | Limitation |
|---|---|---|
| Development | 410 | Includes narrow questions about components |
| Editorial policy | 160 | The team’s most successful category |
| Design | 93 | Answers currently cover only atomic components |
| Total | 663 | Almost 60% received a specific answer from the documentation |
DAIS development plans
The team plans to enable work with layouts through MCP, place the agent in support chats for initial guidance and automate knowledge-base updates. Human support responses will remain in place for now.
Sources
Event date: 2026-09-30. Primary source date: 2026-09-30.
