Skip to main content

​
Questions this guide answers

  • Which settings affect creativity vs consistency?
  • Why are responses short or long?
  • How much context and max tokens can my plan use?
  • What does web search actually do?
  • How do streaming and social state change platform replies?

​
Core controls

​
System instructions (behavior prompt)

System instructions are your verb’s persistent behavior rules. Where to set:
  • Dashboard -> Bot -> AI Engine -> Behavior
  • Field name: systemInstructions
Limit:
  • Up to 8000 characters
How they interact with other layers:
  • systemInstructions: behavior/rules/tone constraints
  • Training examples: style shaping and response pattern hints
  • Long-term memory: durable facts/preferences
  • Knowledge entries: factual/reference content
  • Conversation context: recent turns in current chat/thread/session
Practical rule:
  • Put “how to behave” in system instructions.
  • Put “facts to remember” in knowledge/memory.
  • Put “how to phrase outputs” in training examples.

​
Plan-based limits

Current default limits:
If you request settings above your tier limits, they are clamped or rejected depending on the endpoint.

​
Premium AI model access

AI model availability is plan-based. If you open the AI Engine and choose a model that is outside your current tier, Verba shows an upgrade prompt with the number of additional premium AI models available above your current plan. That upgrade hint is dynamic, so the exact count can differ depending on:
  • Your current plan
  • The active model catalog
  • Whether new premium models have been added since your last visit

Reliable support

Temperature 0.4-0.7, top-p 0.7-0.9, moderate context.

Creative roleplay

Temperature 0.8-1.1, top-p 0.9-1.0, higher context.

Cost-aware

Lower context, web search off, shorter reply style.

Fast troubleshooting

Lower context + deterministic settings to reduce latency variance.

​
Multi-message behavior

When enabled:
  • Verba can split a response into multiple shorter messages.
  • Delay between parts is configurable (0..10000ms).
Use this for natural chat pacing; disable if you want one compact answer.

​
Response streaming

When enabled, Discord, WhatsApp, and Telegram can show the beginning of a response early and edit that message into the complete answer. This is separate from Ultra-only API SSE streaming. Use it for better perceived speed on long answers. Turn it off when your style depends on several intentionally separate messages.

Response streaming

See supported surfaces, fallback behavior, and API differences.

​
Social state

The Social tab adds eight temporary sliders for willingness to talk, irritation, focus, energy, confidence, affection, playfulness, and curiosity. Set them manually or enable Auto mode so recent exchanges can move them gradually. This layer influences delivery without replacing personality, knowledge, memory, or safety rules.

Social and emotional state

Learn the ranges, auto-mode behavior, and memory interaction.

​
Web search behavior

When enabled, Verba may perform model-driven search planning before the final answer. This improves freshness, but can increase:
  • Latency
  • Token usage
  • Cost
Use web search for:
  • News
  • Live pricing
  • Fast-changing product details
Keep it off for:
  • Roleplay
  • Stable canon/lore
  • Deterministic support flows

​
System instructions best practices

  • Keep instructions concrete and scoped.
  • Avoid contradictory rules.
  • Prefer short imperative bullets over long prose.
  • Include failure policy (what to do when unknown).
  • Specify output format explicitly (headings, bullets, steps, code blocks).

​
Quick diagnostics

Lower temperature, tighten system instructions, and add targeted training examples.
Increase model context (within your plan limit).
Increase max tokens and use a fuller reply style.
Reduce context and disable web search unless needed. Response Streaming can make the reply appear sooner, but does not reduce total generation time.

Bot Settings Reference

See all AI-related fields and limits in one place.