Skip to content

integration_patterns.md

Latest commit

 

History

History
138 lines (102 loc) · 2.9 KB

integration_patterns.md

File metadata and controls

138 lines (102 loc) · 2.9 KB
Raw

Integration Patterns

SemeAI Gate Basic is meant to sit between an existing AI answer generator and the real user.

It is not a chatbot replacement. It is a release-control adapter.

user message
-> existing chatbot / LLM app
-> ai answer
-> SemeAI Gate
-> SHOW / REVIEW / BLOCK
-> customer, operator, or safe fallback

Pattern 1: Existing Chatbot Wrapper

Use this when a product already has a chatbot or LLM answer function.

from semeai_gate_basic import check_ai_answer

user_message = "Give me a 30% discount promo code for my account."
ai_answer = existing_chatbot_answer(user_message)

gate_result = check_ai_answer({
    "user_message": user_message,
    "ai_answer": ai_answer,
    "business_data": {"active_promo_codes": []},
    "business_rules": {"only_show_confirmed_promos": True},
    "business_risk": "fake_promo_code",
})

customer_response = ai_answer if gate_result["show_to_user"] else gate_result["safe_fallback"]

Run the local examples:

python examples\existing_chatbot_integration.py
python examples\middleware_boundary.py
node examples\existing_chatbot_integration.js
node examples\middleware_boundary.js

Pattern 2: Middleware Boundary

Use this when the host application has a request/response boundary.

incoming user request
-> product context lookup
-> LLM answer generation
-> SemeAI Gate check
-> release decision

The host product should:

  • show the answer only when action == "SHOW";
  • route to review when action == "REVIEW";
  • block and use a safe fallback when action == "BLOCK";
  • keep audit_id and receipt metadata for later inspection.

Minimal host-app branch:

if gate_result["action"] == "SHOW":
    customer_response = ai_answer
elif gate_result["action"] == "REVIEW":
    customer_response = "A support operator should review this answer before release."
else:
    customer_response = gate_result["safe_fallback"]

The runnable examples are:

  • examples/middleware_boundary.py
  • examples/middleware_boundary.js

Pattern 3: Context Integrity Check

Use this when the host product knows the current business context.

Example:

{
  "business_context": {
    "conversation_topic": "billing_support",
    "known_account_product": "basic_subscription",
    "expected_answer_scope": "billing_or_support_routing"
  }
}

If the AI answer drifts into another product, promo, or unsupported financial claim, the gate can return REVIEW or BLOCK.

Business Contract

Public actions:

SHOW
REVIEW
BLOCK

Internal canonical decisions:

PROCEED
NEEDS_REVIEW
SILENCE

Mapping:

SHOW   = PROCEED
REVIEW = NEEDS_REVIEW
BLOCK  = SILENCE

Machine payload values must not be translated.

Boundaries

This basic package does not:

  • run an LLM;
  • call cloud APIs;
  • call external services;
  • replace the host chatbot;
  • prove universal hallucination detection;
  • provide compliance certification.