Skip to content

armgit / laya

Public Imported from GitHub
0

Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.

fix(integrations): gate a LayaTaskGuard score question on its upper-half probability

abhijithneilabraham0a06f6622299
diff --git a/laya/integrations/_guard.py b/laya/integrations/_guard.py
new file mode 100644
index 00000000..c8d482e9
--- /dev/null
+++ b/laya/integrations/_guard.py
@@ -0,0 +1,19 @@
+"""The score-question rule shared by the guardrail wrappers."""
+from typing import Any, Dict
+
+
+def score_violation_probability(answer: Dict[str, Any], levels: int = 0) -> float:
+ """Probability that a `score` answer is at or above the middle of its scale.
+
+ `score` is the expected level (0..k-1), not a probability, so a guard threshold cannot be
+ compared with it directly. `levels` is the number of levels the question defines; it is only
+ used when the answer carries no `probabilities`, and then the normalised expected level stands
+ in for the distribution.
+ """
+ probs = answer.get("probabilities") or {}
+ k = len(probs) or levels
+ if k < 2:
+ return 0.0
+ if probs:
+ return sum(float(probs.get(str(i), 0.0)) for i in range(k // 2, k))
+ return answer.get("score", 0.0) / (k - 1)
diff --git a/laya/integrations/crewai.py b/laya/integrations/crewai.py
index 057349d4..4d045e54 100644
--- a/laya/integrations/crewai.py
+++ b/laya/integrations/crewai.py
@@ -57,6 +57,7 @@ class LayaTaskGuardError(ValueError):
from ._controls import budget_kwargs as _budget_kwargs, hook_kwargs as _hook_kwargs # noqa: E402
from ._controls import decision_kwargs as _decision_kwargs # noqa: E402
from ._controls import predict_kwargs as _predict_kwargs, reject_remote_hooks as _reject_remote_hooks # noqa: E402
+from ._guard import score_violation_probability as _score_violation_probability # noqa: E402
from ..confidence import _gate_confidence # noqa: E402
# One class for every integration, so `except LayaLowConfidenceError` catches all of them.
@@ -448,6 +449,8 @@ class LayaTaskGuard:
hooks_raise: Optional[bool] = None,
hooks_timeout: Optional[float] = None,
):
+ if not 0.0 <= threshold <= 1.0:
+ raise ValueError("threshold must be a probability in [0, 1]; got %r" % (threshold,))
self.questions = questions
self.action = action
self.rejection_message = rejection_message
@@ -504,11 +507,17 @@ class LayaTaskGuard:
"probability": ans["noul"],
"confidence": ans.get("confidence", 0.0),
}
- elif t == "score" and ans.get("score", 0.0) >= self.threshold:
- violations[qid] = {
- "score": ans["score"],
- "confidence": ans.get("confidence", 0.0),
- }
+ elif t == "score":
+ # `score` is the expected level (0..k-1), not a probability: gate on the
+ # probability that the level is at or above the middle of the scale.
+ levels = len(qdefs.get(qid, {}).get("criteria") or [])
+ p_violation = _score_violation_probability(ans, levels)
+ if p_violation >= self.threshold:
+ violations[qid] = {
+ "score": ans.get("score", 0.0),
+ "probability": round(p_violation, 4),
+ "confidence": ans.get("confidence", 0.0),
+ }
is_safe = len(violations) == 0
diff --git a/laya/integrations/langchain.py b/laya/integrations/langchain.py
index 3f60dbf1..87386a0a 100644
--- a/laya/integrations/langchain.py
+++ b/laya/integrations/langchain.py
@@ -34,6 +34,7 @@ from ._controls import budget_kwargs as _budget_kwargs, hook_kwargs as _hook_kwa
from ._controls import decision_kwargs as _decision_kwargs
from ._controls import predict_kwargs as _predict_kwargs
from ._controls import reject_remote_hooks as _reject_remote_hooks
+from ._guard import score_violation_probability as _score_violation_probability
class LayaGuardrailError(ValueError):
@@ -654,16 +655,9 @@ class LayaGuardrail(_BatchedRunnable, RunnableSerializable):
}
elif t == "score":
# `score` is the expected level (0..k-1), not a probability: gate on the
- # probability that the level is at or above the middle of the scale. Without
- # a distribution, the normalised expected level stands in for it.
- probs = ans.get("probabilities") or {}
- k = len(probs) or len(self._questions().get(qid, {}).get("criteria") or [])
- if k < 2:
- p_violation = 0.0
- elif probs:
- p_violation = sum(float(probs.get(str(i), 0.0)) for i in range(k // 2, k))
- else:
- p_violation = ans.get("score", 0.0) / (k - 1)
+ # probability that the level is at or above the middle of the scale.
+ levels = len(self._questions().get(qid, {}).get("criteria") or [])
+ p_violation = _score_violation_probability(ans, levels)
if p_violation >= self.threshold:
violations[qid] = {
"score": ans.get("score", 0.0),