Guardrails & AI Behavior

J.O.S.I.E is designed to be helpful but bounded — not a general-purpose chatbot.

Pipeline overview

User message
    ↓
Content filter (off-topic / unsafe signals)
    ↓
RAG retrieval (Pro+ only, if corpus exists)
    ↓
Context guard (require evidence for factual claims)
    ↓
LLM with system prompt + static knowledge + RAG context
    ↓
Response to user

Content filter

Before any model call, server-side checks evaluate the message for:

• Off-topic requests relative to configured scope
• Prompt injection patterns
• Disallowed categories (configurable via prompt)

Blocked requests return a polite refusal without calling the LLM.

Context guard

For factual questions (pricing, policies, product specs):

• If RAG chunks or static knowledge cover the topic → answer allowed
• If no evidence → assistant explains it doesn't have that information
• Prevents inventing refund windows, prices, or features

This is why good documentation matters on Pro plans.

RAG grounding

When chunks are retrieved:

• Context is injected with clear boundaries
• Model instructed to prefer provided text
• Distance thresholds vary by tier (Pro/Enterprise are more permissive for FAQ phrasing)

Demo vs. live behavior

Behavioral guardrails are the same; model quality differs.

Tuning behavior without code

1. Tighten system prompt — explicit "never discuss X"
2. Add static knowledge for must-not-get-wrong facts
3. Upload RAG docs for detailed catalogs
4. Test refusal paths — "write me a poem", competitor questions

What J.O.S.I.E is not

• Not a replacement for legal/medical/financial advice
• Not trained on your data globally — only your tenant config + docs
• Not guaranteed 100% accurate without proper knowledge coverage

Human escalation paths should stay in your prompt and static knowledge.