The rogpOS assistant was designed to operate on the device, without depending on the cloud for normal response. The priority is to find short and useful local evidence. The model comes in later, when the local base does not resolve the question.
Current state: The default model is Gemma 3 270M IT Q8_0, run locally. There is no active integration with the Gemini API. Gemini and Gemma are different products; future integration with Gemini would require own product decision, connectivity, data policy and consent.
The current flow
- The question is standardized and associated, when possible, with the game or open guide.
- The local RAG searches the knowledge package for relevant records.
- Results are ranked and limited to a small set of evidence.
- If the local evidence already responds, the system returns a literal source response, without starting the model.
- When it is not enough, the local runtime reserves Gemma, sets up a limited prompt and generates the response.
- The response remains subject to the evidentiary threshold and may declare that it has not found sufficient basis.
The runtime uses a local Unix socket, a queue with a worker, limited buffers, and cancellation. During gameplay, the AI can be paused to return memory and CPU to the main path.
What does RAG mean here
RAG is recovery increased per generation. In current rogpOS, recovery is local and driven by records and keywords. It does not depend on embeddings or a remote vector database. For game guides, the design provides for few relevant excerpts and citation chips, with literal fallback when the template is not available.
Embeddings, like those offered by the Gemini API, can represent text for semantic search. This is a future architectural possibility, not an active feature of rogpOS.
Limits that need to appear
- a small model may misinterpret context, names, and intent;
- the response doesn't know anything that isn't in the local package or the prompt;
- keyword retrieval misses semantic relations and unexpected synonyms;
- Community content may contain errors and needs provenance;
- little memory requires short context and few results;
- a fluent response is not proof that the instruction is safe;
- the wizard must not invent missing compatibility, commands, or facts.
Privacy and network
The default path doesn't send the question to a cloud AI service. If remote integration is studied in the future, it should be turned off by default, explain what data leaves the device, ask for specific consent, offer a local alternative, and allow revocation.
How to improve without losing control
- version knowledge packages;
- record the origin, scope and date of each section;
- assess recovery and response separately;
- maintain a set of known questions and measure accuracy, denial, and citation;
- compare lexical and semantic search with the same corpus before adding infrastructure;
- prevent the model from executing commands directly;
- reserve CPU and memory for the gaming experience.
The goal is not to answer everything. It’s about helping with clarity and knowing when there isn’t enough evidence.