Hallucination (LLM)

A language model predicts the most plausible next words given everything before them. When the true answer is in its context, plausible and correct usually coincide. When it is not, when you ask about your clinic's Friday hours and the model never saw them, the most plausible continuation is still a confident sentence with a time in it. The model has no internal flag for 'I do not know this'; it has to be given one through the way the system is built.

For a business, a hallucinated answer is a liability, not an inconvenience. An agent that invents a discount, misstates a return window, or tells a patient a medication dose creates a refund, a complaint, or worse. Arabic adds a specific risk: models trained mostly on English and Modern Standard Arabic are more likely to fill gaps when the question arrives in Gulf dialect or Arabizi, because the input itself is less familiar. That is why 'it works in the demo' is not evidence; the demo asked questions the vendor had prepared.

Reducing hallucination is an engineering discipline with three layers. Retrieval-augmented generation (RAG) puts the business's real documents in front of the model so the answer is grounded in text it can quote. Guardrails constrain what the model may say, force it to decline or hand off when retrieval returns nothing, and block specific claims (prices, medical advice) unless they come from a trusted source. Golden-set evals measure the remaining error rate on real customer questions before every release. Nano AI ships all three with every agent and reports the eval results to the client; that is what the 30-day measured report in the Launch plan documents.