Spaces:
Running on Zero
Running on Zero
File size: 3,824 Bytes
51da1ee cbe9a94 51da1ee cbe9a94 51da1ee cbe9a94 51da1ee cbe9a94 51da1ee cbe9a94 51da1ee cbe9a94 51da1ee cbe9a94 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | # 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),
}
|