Call centers of government departments are overwhelmed with repetitive questions: how to get a certificate, what documents are needed, where is the application status. Each conversation costs the budget 300–500 rubles, and waiting time reaches 15 minutes. We develop AI chatbots that respond instantly, around the clock, with accuracy down to the latest regulatory revision. Our experience — 20+ projects in the public sector, over 5 years in the market. Average savings per contact are substantial, and call center deflection rate reaches 40–60%. An AI chatbot for government services automates citizen information, reducing the load on the call center.
Specifics of a government bot
Accuracy is critical: information about document requirements, deadlines, procedures must be correct. RAG (Retrieval-Augmented Generation) on current regulatory documents with regular database updates when legislation changes. We use high-dimensional embeddings (768+ dim) and re-ranking to eliminate hallucinations.
Accessibility: the bot works for different categories of citizens — elderly (simple language), foreigners (multilingual via LLM), people with disabilities (voice mode, screen reader). Voice input support based on Whisper.
Neutrality: a neural network chatbot for the government does not express opinions on political issues, does not recommend commercial services. Answers strictly follow procedures, with a disclaimer about relevance.
How to ensure answer accuracy in a government AI bot?
We use RAG with a vectorized document database (decrees, orders, regulations). On request, top-5 fragments are extracted, context truncated to 4000 tokens. LLM (GPT-4o, LLaMA 3 70B) generates an answer only based on these fragments — this reduces hallucinations to 2%. Regular database update via a parser with GitHub (tracking changeset). Each answer contains an update date and source: "Information current as of [date]. Check requirements at [agency].ru". According to recommendations of the Ministry of Digital Development, this approach meets requirements for informing citizens.
| Approach | Accuracy | Latency (p95) | Maintenance complexity |
|---|---|---|---|
| RAG + LLM | 95%+ | 2–4 s | Low (update index) |
| Fine-tune LLM | 85% | 0.8 s | High (retrain on changes) |
| Rules + regex | 70% | 0.1 s | Medium (manual rule editing) |
Typical scenarios and response time
| Query type | Average response time | Accuracy |
|---|---|---|
| Application status | 0.5 s | 98% |
| Documents for service | 2 s | 95% |
| Step-by-step instruction | 3 s | 93% |
- "What documents are needed for registration at the place of residence?" — exact list with link to current regulation
- "How to get a foreign passport through Gosuslugi?" — step-by-step instruction with ESIA authorization
- "Where to get a certificate of no criminal record?" — agency, methods, deadlines
- "Status of my application" — integration with Gosuslugi API
- "How to apply for benefits?" — document list, review times, online submission
Integration with EPGU (Gosuslugi)
REST API of Gosuslugi allows checking application status, informing about necessary actions. Authorization via ESIA is the only way to get personalized information. For calls we use OAuth 2.0 with client credentials. We guarantee 99.9% uptime through fault-tolerant deployment on Kubernetes. As specified in the ESIA API specification, such authorization meets security requirements.
Why RAG is preferable to fine-tuning for regulatory documents?
Fine-tuning memorizes data distribution, but when one order changes, retraining is required. RAG fetches current documents dynamically — just upload a new version to the vector database. This speeds up updates from weeks to minutes. RAG surpasses fine-tuning in update speed by 5 times and in cost by 4 times. Savings on retraining — up to 80%.
When RAG fails: fallback strategies
When answer confidence is low (score < 0.7), the bot activates fallback scenarios: asks clarifying questions, suggests rephrasing the query, or switches to an operator. Log monitoring allows monthly model retraining on new questions. This hybrid architecture maintains accuracy >95% even with increasing query volume.
Example scenario: getting a foreign passport
The user asks: "How to get a foreign passport through Gosuslugi?". The bot extracts the current administrative regulation from the vector database, forms a step-by-step instruction: authorize in ESIA → fill application → upload photo → pay fee → visit MFC. If the user is not authorized, the bot offers to authenticate.
Work process
- Analytics: audit your regulations, identify top-50 questions from call center logs.
- Design: choose LLM for the government sector, vector database, scenarios (authorization, fallback to operator).
- Development: implement RAG pipeline for government agencies, integrate with ESIA/SMEV, set up voice channel.
- Testing: expert audit of 1000 random responses (accuracy >95%), load testing (100 rps).
- Deployment and support: deploy in your environment, monitor, retrain monthly.
What is included in the work
- Documentation: API specification (OpenAPI), operation manual
- Code: Git repository, CI/CD pipeline
- Access: to vector database, logs, monitoring dashboard
- Training: workshop for operators on configuration and retraining
- Support: 3 months included (SLAs: 4 hours for critical errors)
Timeline and cost
Timeline — from 4 weeks (basic bot) to 12 weeks (full-featured). Cost is calculated individually after reviewing your scenarios. Contact us for a free assessment of your project. Average call center deflection rate reaches 40–60%, allowing up to 80% savings on citizen support budget. Get a consultation on implementation — estimate the effect for your agency.