Send a message to an assistant

Variations are almost always caused by one (or several) of these factors:

• High temperature: at higher values (≈ 0.8-1.0) the model injects more randomness. Lower it to 0.2-0.4 if you need consistency.
• Missing examples or clear rules: without few-shot examples or precise instructions, the model fills in gaps on its own. Add concrete input-output samples or define rules in the system prompt/instruction.
• Ambiguous prompt: if your text supports multiple interpretations, the model will choose one at random. Rephrase so there is only one possible reading.
• Context shifts: any previous message in the conversation influences the answer. Restart the thread or explicitly summarise the context if you need reproducibility.
• Token limits: when the conversation grows long, parts of the history may be truncated and the model "forgets" details. Shorten the prompt or summarise the context.

Follow these recommendations to obtain much more stable and predictable responses.