AI tool disclosure

Models used

OpenAI gpt-4o for assessment-item personalisation, the in-product assistant, training-exercise generation, and the rationale that ties each recommendation to its retrieved source passages.

Cohere embed-v4.0 (multilingual, 1024-dim) for the retrieval-augmented knowledge base, paired with Cohere rerank-v3.5 as a second-stage multilingual reranker. Cohere is used here, rather than OpenAI embeddings, because it gives German and English queries comparable recall on the bilingual corpus.

Where AI is used

• Assessment items are rephrased per role and tasks; the underlying validated MAILS wording is preserved.
• Recommendation evidence is retrieved from a curated knowledge base, ranked by cosine similarity, and shown with a confidence band.
• Training items are generated from canonical templates and personalised by the LLM under per-type integrity constraints.

Graceful degradation

If the LLM is unreachable or slow, the product opens a 60-second process-wide circuit breaker.

While the breaker is open, every AI feature falls back to verbatim validated content. No screen is blocked.

What is sent to the AI - and what is not

Only the data needed for the immediate task leaves the product. OpenAI receives the item being personalised plus a minimal role context (job role, domain, primary tasks). Cohere receives only search-query text and the curated knowledge-base passages that are being indexed or re-ranked. Your raw Likert responses never leave Cloud SQL.

We do not train any external model on your data. AI calls travel over HTTPS to OpenAI (US-hosted endpoints) and Cohere (US- and EU-hosted endpoints); data at rest stays in Cloud SQL in Switzerland (europe-west6). Both providers' API terms prohibit training their general-purpose models on customer prompts and inputs.