Configuration
Temperature
Tune temperature (0.0 to 1.0) and top-p sampling to balance factual grounding with conversational naturalness.
Overview
Temperature governs the randomness of token selection during LLM inference. Lower values make completions deterministic and strictly aligned with provided documentation; higher values introduce stylistic variation.How Temperature Affects AI Responses
- Temperature 0.0 – 0.2 (Strict / Factual): The model consistently chooses the highest-probability token. Essential for technical documentation, pricing disclosures, and legal policies.
- Temperature 0.3 – 0.6 (Balanced): Natural phrasing with high factual grounding. Recommended default for product onboarding.
- Temperature 0.7 – 1.0 (Creative / Exploratory): Varied vocabulary and open-ended ideation. Used for brainstorming assistants.
Recommended Temperature by Use Case
| Use Case | Recommended Temperature | Top-P | Rationale |
|---|---|---|---|
| API & Technical Documentation | 0.1 | 0.9 | Strict adherence to code syntax |
| E-Commerce & Product Guide | 0.3 | 0.9 | Factual specs with friendly tone |
| Customer Support FAQ | 0.2 | 0.85 | Consistent answers to recurring questions |
| Sales & Marketing Concierge | 0.5 | 0.95 | Engaging conversational flow |
Top-P (Nucleus) Sampling
In addition to temperature, Asiyst supports Top-P sampling (default: 0.9), which restricts candidate tokens to the cumulative probability mass $P$, pruning low-probability outliers.