A controlled intelligence layer for orders.
Orderix.ai connects domain training, retrieval, AI decisions, deterministic controls and feedback.
MEDVION
Order
Input order data.
RAG
Retrieve context.
LLM
Generate decision.
Guardrails
Validate authority.
Agent
Execute or escalate.
Feedback
Log and retrain.
RAG
Historical cases become context.
The POC uses Sentence Transformers and pgvector to retrieve relevant order and resolution records before the AI generates a decision.
Vector embeddings
Order history and resolution records use vector(384) fields.
Similarity search
The example retriever returns the top three matching cases.
Prompt context
Retrieved cases are inserted into the agent prompt.
Safety
Five deterministic guardrails.
Input validation, action whitelist, business rules, confidence and action-tier controls.
Input validation
Checks order_id, items and total.
Action whitelist
Blocks invalid actions.
Business rules
Applies thresholds such as high-value refund review.
Confidence
Blocks decisions below the POC threshold.
Tier control
Maps actions to automatic or human-review paths.
Authority boundary
The AI never decides its own authority level.
