Responsible AI

OPSQAI is designed around a single principle: an AI answer is only useful if it can be verified. The product ships as a self-hosted Windows application; the AI provider is chosen and operated by the customer.

Source grounding

OPSQAI answers strictly from documents that have been retrieved from the customer's own knowledge base by the local RAG worker. When no source supports an answer, the model is instructed to refuse rather than to guess.

Citations

Every answer includes the source document, section and a verbatim excerpt, so end users and auditors can verify the answer against the original SOP.

Model providers — chosen by the customer

Inference is performed by the AI provider the customer configures during the first-run wizard. Currently supported adapters:

• OpenAI — customer's own API key.
• Azure OpenAI — customer's own tenant, endpoint and deployment.
• Ollama — fully local inference, no external calls.
• OpenRouter or any OpenAI-compatible endpoint.

OPSQAI itself is not an AI provider and does not proxy AI traffic. Any commitment about training on customer content (e.g. OpenAI's "no training on API data" default, or Azure OpenAI's contractual terms) applies through the customer's own contract with the chosen provider — not through OPSQAI.

Local-only option

Customers with strict data-residency requirements can run OPSQAI end-to-end on premises by selecting the Ollama adapter: embeddings, retrieval and generation stay on the customer's Windows Server, with no outbound AI traffic.

Human oversight

OPSQAI surfaces an internal request workflow when the AI cannot answer. Managers review these gaps and decide whether to publish an SOP or FAQ. The audit log records every question, retrieval and answer for later review by administrators.

Safety-critical use

OPSQAI is a decision-support tool. Customers must review AI output before acting on it in safety-critical contexts (forklift operation, hazardous goods, emergency procedures).

Bias and language

OPSQAI supports English, German and Romanian. Outputs in any language are translated from the same underlying sources to keep meaning consistent. We monitor for translation drift and welcome reports at notify@opsqai.de.