AI configuration

Written by Aashiq, Founder, AskVault · Reviewed by Aashiq

Last updated: May 15, 2026 · 4 min read

What AI Config controls

Six knobs that shape every bot response:

System prompt. The personality, tone, and rules layered on top of every conversation.
Response length. Max characters per bot reply.
Citation policy. How and when the bot cites sources.
Fallback message. What the bot says when it can't answer.
Language detection. Auto-detect vs forced single language.
Off-topic handling. What to do with unrelated questions.

All under Settings > AI Config.

System prompt

The most impactful knob. A system prompt of 200 to 500 words shapes how the bot behaves across thousands of conversations.

Default prompt:

You are a helpful assistant for {workspace_name}. Answer based on the provided knowledge. If you don't know, say so. Cite sources when relevant. Be concise.

Customizable to:

Match your brand voice. Formal, casual, technical, friendly.
Set the bot's role. "You are a customer support agent. Be empathetic."
Define escalation triggers. "If the user expresses frustration, suggest escalating to a human."
Specify off-limits topics. "Don't discuss competitor pricing."

Edit under Settings > AI Config > System Prompt. Test in the Chat Playground before publishing.

System prompt patterns

Three patterns that work for most B2B SaaS:

Pattern 1: support agent.

You are a knowledgeable customer support agent for {workspace_name}. Answer questions from the knowledge base. Be concise (under 100 words). Cite the source when relevant. If the user is frustrated or asks for a human, suggest they say "talk to a human". Don't promise refunds or discounts beyond stated policy.

Pattern 2: sales engineer.

You are a sales engineer for {workspace_name}. Help prospects understand our product, pricing, and integrations. Be enthusiastic but accurate. If asked about pricing for Enterprise, suggest booking a demo. Don't reveal internal roadmap details.

Pattern 3: internal helpdesk.

You are an HR helpdesk for {company}. Answer employee questions about benefits, PTO, payroll. Be empathetic and clear. Cite policy documents. For sensitive issues (harassment, legal), suggest contacting HR directly via email.

Adapt to your context. Tone tweaks compound across thousands of interactions.

Response length

Three modes:

Short. Under 80 characters (good for SMS).
Medium. Up to 300 characters (good for chat).
Long. Up to 1,000 characters (good for email or detailed support).

Auto-mode picks based on channel (SMS = short, email = long).

Override per skill if needed.

Citation policy

How the bot cites:

Always. Every claim cites a source. Best for compliance-sensitive use cases.
When asked. Bot cites only when the user asks "where did you get this?".
Never. No citations. Rarely a good choice; users distrust uncited claims.

Default: always. Citations are the main trust mechanism in RAG-based bots.

Fallback message

When the bot can't answer (knowledge gap):

"I don't have specific information about that. Want me to connect you with our team?"

Customize for your tone. Some teams prefer humor, some technical, some apologetic.

Trigger conditions:

No matching retrieval chunks (similarity below threshold).
Question explicitly out of scope.
Question too vague to answer.

The bot logs each fallback as a knowledge gap. Review under Analytics > Unanswered Questions.

Language detection

Three modes:

Auto. Detect from each visitor message. Reply in the same language. Switches mid-conversation if visitor switches.
Workspace default. Always reply in workspace's default (e.g., always English).
Per-channel override. Different defaults per channel (e.g., WhatsApp in Spanish, widget in English).

35 languages supported.

Off-topic handling

How the bot handles unrelated questions:

Polite redirect. "I can only help with {workspace_name}-related questions. Want me to find something for you?"
Strict refusal. "That's outside my scope. Try our knowledge base or contact support."
Best-effort answer. Try to answer generally, with a disclaimer.

Default: polite redirect.

Useful for staying on-topic in marketing-site bots that customers sometimes test with random questions.

Temperature (advanced)

For Business and Enterprise:

Low temperature (0.0 to 0.3). Conservative, deterministic responses. Same question gets the same answer.
Medium (0.4 to 0.7). Balanced. Slight variation in wording.
High (0.8 to 1.0). Creative, varied. Higher hallucination risk.

Default: medium-low (0.3). Most teams don't change this.

Lower temperature is safer for fact-heavy use cases (support, billing). Higher is for creative use cases (suggestion engines, marketing).

Model routing

By plan tier (set automatically):

Free. Lightweight model. Fast, capable.
Starter, Growth. Mid-tier model. Better at reasoning.
Business. Higher-capability model. Best for complex queries.
Enterprise. Custom model selection. Bring your own model.

Override under Settings > Model Selector (Business and above).

Custom instructions per skill

Some skills override the workspace system prompt:

escalate_to_human uses a more empathetic tone.
collect_lead uses a more inviting tone.
subscription_manager uses a more formal tone (regulatory caution).

Customize each under AI Agents > Skills > [skill] > Messages.

Multi-language brand voice

For workspaces serving multiple languages:

Define language-specific prompt overrides under AI Config > Multi-Language.
Bot uses the localized prompt based on detected language.
Default prompt in English fallback for unsupported languages.

Useful for global teams with regional tone preferences.

A/B testing prompts

For Growth and above:

AI Config > Experiments > Create.
Pick a variant prompt.
Set traffic split (e.g., 50/50).
Define success metric (CSAT, resolution rate, conversion).
Run for at least 2 weeks.
Promote the winner.

Useful for incremental prompt optimization.

Prompt template variables

Use template variables to inject context:

\{workspace_name\}. Your workspace name.
\{visitor_name\}. Visitor's first name (identity-verified only).
\{visitor_plan\}. Visitor's plan tier (identity-verified only).
\{current_date\}. Today's date.
\{channel\}. Current channel.

The bot resolves these at conversation start.

Audit and rollback

Every prompt edit logs:

Old and new prompt (full diff).
Author and timestamp.
Effective from (when change took effect).

Rollback under AI Config > History > Rollback. Available for 90 days post-change.

API access

For programmatic config:

curl -X PATCH https://api.askvault.co/v1/ai-config \
  -H "Authorization: Bearer ak_xxx" \
  -H "Content-Type: application/json" \
  -d '{"system_prompt": "...", "response_length": "medium", "temperature": 0.3}'

Useful for version-controlling prompts in git.

Planned features (on the roadmap)

Documented for accuracy:

Visual prompt editor. Today, raw text. Block-based visual editor planned.
Prompt linting. Today, free-form. Pre-deploy lint for common issues planned.
Multi-step reasoning prompts. Today, single-turn. Chain-of-thought planned for Enterprise.
Per-audience prompts. Today, one prompt per workspace. Per-audience prompts (different tone for free vs enterprise) planned.

Limits

System prompt length. 4,000 characters.
Response length. Up to 1,000 characters per reply.
A/B variants. Up to 4 simultaneous.
History retention. 90 days of prompt versions.

Common pitfalls

Bot too verbose. Response length too high. Reduce to medium or short.

Bot too generic. System prompt vanilla. Add specific brand voice, role, and constraints.

Bot hallucinates. Temperature too high. Lower to 0.2 or 0.3.

Bot doesn't escalate when frustrated. Trigger condition not in prompt. Add explicit escalation cue.

Bot citations breaking. Citation format conflicts with channel rendering. Test in each channel; SMS needs plain-text citations.

FAQ

Will changing the system prompt affect ongoing conversations?

Existing conversations continue with the old prompt. New conversations use the new prompt.

Can I write prompts in non-English?

Yes. The bot understands prompts in any supported language and responds in matched language.

How long should my system prompt be?

200 to 500 words is typical. Longer than 1,000 words diminishes returns and uses LLM token budget.

Can I test prompts without deploying?

Yes via the Chat Playground.

Does the bot follow my prompt strictly?

Mostly. AI models can occasionally deviate. For strict enforcement (regulatory caution), use audience tags and skill policy bounds.

Was this page helpful?