The generative approach creates chatbot responses dynamically from user messages instead of selecting them from a predefined answer set. It offers flexibility for edge cases, but makes output quality harder to control and predict. section 12.4
A practical generative system can combine:
- A pretrained encoder–decoder network, typically a transformer, or a larger language model. section 12.4
- A prompt that directs the model’s behavior. section 12.4
- Grounding, fine-tuning, guardrails, and intent recognition to reduce unhelpful, incorrect, or toxic replies. section 12.4
- Intent or sentiment evaluation to reject weak responses, regenerate alternatives, or adjust temperature. section 12.4
