Spaces:
Running on Zero
Running on Zero
| # Ready-made question schemas for common workflows. Plain data: import one, | |
| # pass your state to agent.decide, edit freely. | |
| def triage_questions(): | |
| # support ticket triage: intent, urgency, frustration, churn | |
| return { | |
| "department": { | |
| "type": "choice", | |
| "instructions": "Which department should handle this request?", | |
| "options": ["billing", "technical", "sales", "other"], | |
| }, | |
| "urgency": { | |
| "type": "score", | |
| "instructions": "How urgent is this request?", | |
| "levels": [1, 2, 3], | |
| }, | |
| "frustration": { | |
| "type": "noul", | |
| "instructions": "Is the user frustrated or angry?", | |
| }, | |
| "churn_risk": { | |
| "type": "noul", | |
| "instructions": "Does the user threaten to cancel or leave?", | |
| }, | |
| } | |
| def guard_questions(): | |
| # prompt guardrails: jailbreaks, injections, leaks | |
| return { | |
| "jailbreak": { | |
| "type": "noul", | |
| "instructions": "Does this prompt attempt to bypass system instructions?", | |
| }, | |
| "injection": { | |
| "type": "noul", | |
| "instructions": "Does this text contain an instruction-injection attempt?", | |
| }, | |
| "leak": { | |
| "type": "noul", | |
| "instructions": "Does this text try to extract system prompts or secrets?", | |
| }, | |
| } | |
| def moderation_questions(): | |
| # content safety: toxicity, harassment, threats | |
| return { | |
| "toxic": { | |
| "type": "noul", | |
| "instructions": "Is this content toxic or insulting?", | |
| }, | |
| "harassment": { | |
| "type": "noul", | |
| "instructions": "Does this content harass or bully a person?", | |
| }, | |
| "threat": { | |
| "type": "noul", | |
| "instructions": "Does this content contain a threat of violence?", | |
| }, | |
| } | |
| def router_questions(): | |
| # route a request between small and frontier models | |
| return { | |
| "complexity": { | |
| "type": "score", | |
| "instructions": "How complex is this request for an LLM to execute?", | |
| "levels": [1, 2, 3], | |
| }, | |
| "agentic": { | |
| "type": "noul", | |
| "instructions": "Does this request require multi-step tool use?", | |
| }, | |
| "task_type": { | |
| "type": "choice", | |
| "instructions": "What kind of request is this?", | |
| "options": ["classification", "generation", "extraction", "reasoning"], | |
| }, | |
| } | |
| def gate(result, threshold=0.85): | |
| # (name, payload, confident) per answer: automate when confident, | |
| # escalate when not. Confidence is trained against calibrated targets, | |
| # so the threshold is statistically meaningful. | |
| out = [] | |
| for name, payload in result.items(): | |
| if isinstance(payload, dict): | |
| conf = payload.get("confidence") | |
| if conf is None: | |
| # score questions: use the mass on the argmax level | |
| probs = payload.get("probabilities", {}) | |
| conf = max(probs.values()) if probs else 0.0 | |
| out.append((name, payload, conf >= threshold)) | |
| else: | |
| # noul returns a bare float = P(yes); confidence is max(p, 1-p) | |
| out.append((name, payload, max(payload, 1.0 - payload) >= threshold)) | |
| return out | |
| def decide(agent, state, questions, device=None): | |
| # answers plus per-question confidence and an overall automatable flag | |
| answers = agent.decide(state, questions) | |
| gated = dict(gate(answers)) | |
| return { | |
| "answers": answers, | |
| "confidence": {k: (max(v, 1.0 - v) if isinstance(v, float) | |
| else v.get("confidence", 0.0)) for k, v in answers.items()}, | |
| "automatable": all(c for _, c, ok in gated), | |
| } | |