The Ask.Onix project, which began as an educational prototype of a RAG-based expert system, has evolved into an internal AI assistant for the Onix corporate knowledge base.

Employees can ask questions in natural language, receive AI-generated answers, and immediately check them against references to original sources. The project has become a practical example of how we design AI systems around real business knowledge, rather than isolated LLM experiments.
 

Key features of the solution include:

1. Self-hosted within the company, working in an internal loop where only company employees have access.
    
2. In one of the modes, for demonstration, the system can work on free LLM models provided by the OpenRouter service.
   1. These can also be local models, such as: Lamma4 from Meta, Gemma 2 from Google, or gpt-oss-20b from OpenAI.
   2. You can also use a $20 subscription and it will be enough for hundreds of requests per day and for a company of up to 500 people.
       
3. Separate ability to process files of different formats.
    
4. Quick change of LLM model - you can connect your own.
    
5. Built-in scaling mechanism for more data and more powerful models, without changing the code.

The main technical challenge was not the RAG mechanism itself, but eliminating the dependency on a single AI vendor. Initially, Ask relied on DeepSeek cloud models available through Ollama. When the availability of free models, response speed and output quality became unstable, the team introduced OpenRouter and reworked the AI ​​layer around a vendor-independent architecture.

Instead of simply replacing one model with another, we divided the AI ​​stack into three independent components: response generation, image recognition and embedding. This allowed us to maintain the embedding on the existing on-premises Ollama model while moving generation to alternative models. As a result, the entire knowledge base remained compatible and did not need to be re-indexed during the migration.

The result is a more resilient and extensible AI platform, rather than a solution tied to a single LLM. Under current workloads, Ask continues to operate without AI costs or limit exhaustion reports, while the architecture provides an easy path to paid APIs, additional models, and larger usage volumes as needed.

The next step is to add enhanced product analytics and structured user feedback so that we can measure not only the technical robustness, but also the business value that the assistant provides to employees.