Spaces:
Sleeping
Sleeping
File size: 8,510 Bytes
c4fe0a4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 | """Failure mode taxonomy and detection.
Five failure modes, each with a detector that returns
(detected: bool, detail: str).
Every eval run tags each extraction with its failure modes.
The report shows counts, rates, and examples for each mode.
"""
from dataclasses import dataclass
@dataclass
class FailureTag:
"""One detected failure mode on one case."""
mode: str # e.g. "hallucination"
case_id: str
detail: str # human-readable explanation
extraction: dict # the extraction that triggered it
case: dict # the case that was processed
# --- Failure mode definitions ---
FAILURE_MODES = [
"hallucination",
"omission",
"ambiguity",
"overconfidence",
"language_drift",
]
# --- Detectors ---
def detect_hallucination(extraction: dict, case: dict) -> tuple[bool, str]:
"""Recommendation or root cause attribution with no evidence from source text.
Checks: (1) evidence_quotes is empty, or (2) none of the evidence quotes
actually appear in the ticket_text or conversation_snippet.
"""
evidence = extraction.get("evidence_quotes", [])
if not evidence or all(not q.strip() for q in evidence):
return True, "No evidence quotes provided"
# Check if quotes actually appear in the source text
source_text = (
case.get("ticket_text", "")
+ " "
+ case.get("conversation_snippet", "")
+ " "
+ " ".join(case.get("email_thread", []))
).lower()
fabricated = []
for quote in evidence:
quote_clean = quote.strip().lower()
if quote_clean and quote_clean not in source_text:
# Check if at least a substantial substring matches (>= 10 chars)
found_partial = False
if len(quote_clean) >= 10:
for start in range(0, len(quote_clean) - 9):
chunk = quote_clean[start : start + 10]
if chunk in source_text:
found_partial = True
break
if not found_partial:
fabricated.append(quote)
if fabricated:
return True, f"Evidence not found in source: {fabricated[:2]}"
return False, ""
def detect_omission(extraction: dict, case: dict) -> tuple[bool, str]:
"""Clear signal in the source text that the extraction missed.
Heuristic: checks for high-signal keywords in source text that should
have influenced root_cause or risk_level but didn't.
"""
source_text = (
case.get("ticket_text", "") + " " + case.get("conversation_snippet", "")
).lower()
risk_level = extraction.get("risk_level", "low")
root_cause = extraction.get("root_cause_l1", "").lower()
# Urgent signals that should raise risk_level
urgent_signals = ["cancel", "lawsuit", "legal action", "report to", "regulator"]
has_urgent = any(s in source_text for s in urgent_signals)
if has_urgent and risk_level in ("low", "medium"):
return True, f"Urgent signals in text but risk_level={risk_level}"
# Outage/security signals that should affect root_cause
outage_signals = ["outage", "down for", "service unavailable", "cannot access"]
has_outage = any(s in source_text for s in outage_signals)
if has_outage and root_cause not in ("network", "outage", "service", "infrastructure"):
return True, f"Outage signals in text but root_cause={root_cause}"
# Billing signals
billing_signals = ["overcharg", "double charge", "charged twice", "wrong amount", "refund"]
has_billing = any(s in source_text for s in billing_signals)
if has_billing and root_cause not in ("billing", "payment", "pricing"):
return True, f"Billing signals in text but root_cause={root_cause}"
return False, ""
def detect_ambiguity(extraction: dict, case: dict) -> tuple[bool, str]:
"""Case is genuinely ambiguous but extraction doesn't flag uncertainty.
Detected when: ticket_text is very short OR contains conflicting signals,
but confidence is high and review_required is False.
"""
ticket = case.get("ticket_text", "")
confidence = extraction.get("confidence", 0)
review = extraction.get("review_required", False)
# Very short ticket — hard to be confident
if len(ticket.split()) < 8 and confidence > 0.8 and not review:
return True, f"Very short ticket ({len(ticket.split())} words) but confidence={confidence}"
# Ticket has question marks suggesting ambiguity
if ticket.count("?") >= 3 and confidence > 0.8 and not review:
return True, f"Multiple questions in ticket but confidence={confidence}"
return False, ""
def detect_overconfidence(extraction: dict, case: dict) -> tuple[bool, str]:
"""High confidence but wrong root cause (requires gold label).
Also triggers if confidence is very high but risk signals are contradictory.
"""
confidence = extraction.get("confidence", 0)
# Check against gold label if available
gold_root_cause = case.get("gold_root_cause")
if gold_root_cause is None:
# Fallback: check for high confidence with high churn_risk (contradictory)
churn_risk = extraction.get("churn_risk", 0)
risk_level = extraction.get("risk_level", "low")
if confidence > 0.9 and churn_risk > 0.7 and risk_level in ("high", "critical"):
return True, f"Confidence={confidence} but churn_risk={churn_risk}, risk={risk_level}"
return False, ""
predicted = extraction.get("root_cause_l1", "").lower()
gold = gold_root_cause.lower()
if confidence > 0.85 and predicted != gold:
return True, f"Confidence={confidence} but predicted={predicted}, gold={gold}"
return False, ""
def detect_language_drift(extraction: dict, case: dict) -> tuple[bool, str]:
"""Multilingual or format shifts cause classification collapse.
Detected when: case language is non-English or mixed, and the extraction
has low confidence or ambiguous root cause.
"""
language = case.get("language", "en")
confidence = extraction.get("confidence", 0)
root_cause = extraction.get("root_cause_l1", "").lower()
if language in ("mixed", "de", "zh", "unknown"):
if confidence < 0.5:
return True, f"Non-English case (lang={language}) with low confidence={confidence}"
if root_cause in ("unknown", "other", "ambiguous", ""):
return True, f"Non-English case (lang={language}) with ambiguous root_cause={root_cause}"
return False, ""
# --- Main tagger ---
DETECTORS = {
"hallucination": detect_hallucination,
"omission": detect_omission,
"ambiguity": detect_ambiguity,
"overconfidence": detect_overconfidence,
"language_drift": detect_language_drift,
}
def tag_failure_modes(extraction: dict, case: dict) -> list[FailureTag]:
"""Run all failure mode detectors on one extraction.
Returns a list of FailureTag for each detected failure.
"""
tags = []
case_id = case.get("case_id", extraction.get("case_id", "unknown"))
for mode, detector in DETECTORS.items():
detected, detail = detector(extraction, case)
if detected:
tags.append(FailureTag(
mode=mode,
case_id=case_id,
detail=detail,
extraction=extraction,
case=case,
))
return tags
def summarize_failure_modes(all_tags: list[FailureTag]) -> dict:
"""Aggregate failure tags into counts and rates.
Returns:
{
"total_failures": int,
"by_mode": {"hallucination": {"count": N, "examples": [...]}, ...},
"affected_cases": int,
}
"""
from collections import Counter, defaultdict
mode_counts = Counter(t.mode for t in all_tags)
mode_examples: dict[str, list[dict]] = defaultdict(list)
for t in all_tags:
if len(mode_examples[t.mode]) < 3: # Keep up to 3 examples per mode
mode_examples[t.mode].append({
"case_id": t.case_id,
"detail": t.detail,
})
by_mode = {}
for mode in FAILURE_MODES:
by_mode[mode] = {
"count": mode_counts.get(mode, 0),
"examples": mode_examples.get(mode, []),
}
affected_cases = len({t.case_id for t in all_tags})
return {
"total_failures": len(all_tags),
"by_mode": by_mode,
"affected_cases": affected_cases,
}
|