AI system · public sector: standardisation and metrology
RAG knowledge base and AI strategies
Rosstandart · 2025 · AI system · Python 3.11, LangChain
A RAG demo on Rosstandart documents and the “RST 2.0” project, analysis of five subordinate organisations and a specification for extracting requirements from GOST standards
Two related lines of work. The first is a demonstration RAG system on Rosstandart documents and the “RST 2.0” digital transformation project (15 documents, 954 chunks): natural-language answers with links to sources; it went through three iterations (local FAISS, then incremental vector DB management, then migration to Milvus) and a code review. The second is an analytical package on five subordinate organisations (30+ documents, about 600 pages) with AI optimisation strategies, plus a review of a standardisation institute’s technical specification for a system that extracts machine-readable requirements from GOST standards and passes them to the client’s systems.
The task
Scattered information about the agency and its subordinate organisations, and the need for an AI adoption strategy; for the standardisation institute, turning regulatory documents into requirements suitable for machine processing and for integration into AI, MDM and MES systems.
What’s inside
- RAG on Claude 4.5 Sonnet (via OpenRouter) with paraphrase-multilingual-MiniLM-L12-v2 multilingual embeddings; sources cited in every answer
- Edition A: incremental adding and removal of files in the vector DB without full re-indexing, CLI management and integrity validation
- Edition B: Milvus instead of FAISS, protection against prompt injection and DoS, a 60-second timeout on all model requests
- Analytical package for each of the 5 organisations: a comprehensive report, an AI strategy (8 areas, a 2026–2028 roadmap), a short review with SWOT, a work report
- The standardisation institute’s specification: an ontology model (OWL 2, SHACL, Apache Jena), entity and relation extraction (ruRoBERTa-large), a REST API on Java Spring Boot, converters to AI/MDM/MES, an “autonomous start” stage
Machine translation — being proofread.