Ask.Onix.Team started as an educational RAG prototype but quickly evolved into an internal AI assistant for Onix’s corporate knowledge base. The system allows employees to ask questions in natural language, receive AI-generated answers, and immediately see the sources behind them.

The main challenge was not the RAG logic itself, but the dependency on unstable free access to cloud-based models. When the availability and performance of Ollama models deteriorated, the team added OpenRouter and redesigned the AI layer so that answer generation, vision, and embeddings could use different models and providers. The local embedding model was retained, so there was no need to re-index the knowledge base.

The result was a more resilient system that can be switched to an alternative provider within minutes. At the current level of usage, Ask has continued to operate without any AI inference costs, and there have been no reports of usage limits being reached. At the same time, the case highlighted the next priority: adding product analytics and systematic feedback collection to measure not only technical stability, but also the actual value the system delivers to employees.