synth-ai 0.2.13.dev1__py3-none-any.whl → 0.2.14__py3-none-any.whl

synth_ai/task/rubrics/__init__.py

@@ -0,0 +1,56 @@
+"""Rubric schema, loading, and scoring helpers for Task Apps.
+
+This module provides:
+- Flexible rubric models (Criterion, Rubric) for general task app use
+- Strict validators (StrictCriterion, StrictRubric) for step-wise judges
+- Loading utilities supporting JSON, YAML, and HTTP sources
+- Blending utilities for composing rubrics
+- Scoring utilities for events and outcomes
+"""
+
+# Core models (flexible validation)
+from .models import Criterion, Rubric
+
+# Loading and blending
+from .loaders import blend_rubrics, load_rubric
+
+# Scoring
+from .scoring import score_events_against_rubric, score_outcome_against_rubric
+
+# Strict validators (for judge configs)
+from .strict import (
+    StrictCriterion,
+    StrictRubric,
+    ValidationError,
+    validate_rubric_dict,
+    validate_rubric_file,
+    validate_rubric_files,
+)
+
+__all__ = [
+    # Flexible models
+    "Criterion",
+    "Rubric",
+    # Loaders
+    "load_rubric",
+    "blend_rubrics",
+    # Scoring
+    "score_events_against_rubric",
+    "score_outcome_against_rubric",
+    # Strict validators
+    "StrictCriterion",
+    "StrictRubric",
+    "ValidationError",
+    "validate_rubric_dict",
+    "validate_rubric_file",
+    "validate_rubric_files",
+]
+
+# Maintain backwards compatibility
+# Old code may import these names expecting the flexible variants
+RubricCriterion = StrictCriterion
+RubricSpec = StrictRubric
+
+
+
+

synth_ai/task/rubrics/loaders.py

@@ -0,0 +1,152 @@
+"""Rubric loading and blending utilities."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from .models import Criterion, Rubric
+
+
+def _load_text(source: str) -> tuple[str, str | None]:
+    """Load text from file path or return as-is."""
+    path = Path(source)
+    if path.exists():
+        return path.read_text(encoding="utf-8"), path.suffix.lower()
+    return source, None
+
+
+def _parse_structured(text: str, suffix: str | None) -> dict[str, Any]:
+    """Parse JSON or YAML text into a dictionary."""
+    text = text.strip()
+    if not text:
+        raise ValueError("Rubric source is empty")
+    if suffix in (".yaml", ".yml"):
+        try:
+            import yaml  # type: ignore
+        except Exception as exc:  # pragma: no cover - optional dependency
+            raise RuntimeError("PyYAML is required to load YAML rubrics") from exc
+        data = yaml.safe_load(text)
+        if not isinstance(data, dict):
+            raise ValueError("Rubric YAML must produce a mapping") from None
+        return data
+    if text.startswith("{"):
+        return json.loads(text)
+    if text.startswith("http://") or text.startswith("https://"):
+        import requests  # type: ignore
+
+        response = requests.get(text, timeout=15)
+        response.raise_for_status()
+        return _parse_structured(response.text, suffix)
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        try:
+            import yaml  # type: ignore
+        except Exception as exc:  # pragma: no cover - optional dependency
+            raise RuntimeError("PyYAML is required to load rubric text") from exc
+        data = yaml.safe_load(text)
+        if not isinstance(data, dict):
+            raise ValueError("Rubric text must decode to a mapping") from None
+        return data
+
+
+def load_rubric(source: str | dict[str, Any] | Rubric | None) -> Rubric | None:
+    """Load rubric from file path, dict, or return existing Rubric.
+    
+    Args:
+        source: File path (JSON/YAML), dict, existing Rubric, or None
+        
+    Returns:
+        Parsed Rubric instance or None if source is None
+        
+    Raises:
+        ValueError: If the rubric format is incorrect (e.g., backend judge format)
+        ValidationError: If the rubric fails schema validation
+    """
+    if source is None:
+        return None
+    if isinstance(source, Rubric):
+        return source
+    
+    # Load and parse the data
+    if isinstance(source, dict):
+        data = source
+    else:
+        text, suffix = _load_text(str(source))
+        data = _parse_structured(text, suffix)
+    
+    # Check if this looks like a backend judge rubric (wrong format)
+    if isinstance(data, dict) and "event" in data and "outcome" in data:
+        # Missing required task app rubric fields
+        if "version" not in data and "goal_text" not in data and "criteria" not in data:
+            source_hint = f" ({source})" if isinstance(source, str) else ""
+            raise ValueError(
+                f"Rubric appears to be in backend judge format (has 'event'/'outcome' keys){source_hint}. "
+                f"Task apps require rubrics with 'version', 'goal_text', and 'criteria' fields. "
+                f"Backend judge rubrics should be named '*_backend_judge.json' and loaded by judge functions."
+            )
+    
+    return Rubric.model_validate(data)
+
+
+def _merge_weights(base: Criterion, override: Criterion) -> float:
+    """Merge criterion weights from base and override rubrics."""
+    if override.weight != 1.0 and base.weight != 1.0:
+        return base.weight * override.weight
+    if override.weight != 1.0:
+        return override.weight
+    return base.weight
+
+
+def blend_rubrics(base: Rubric | None, override: Rubric | None) -> Rubric | None:
+    """Blend two rubrics by merging criteria and inheriting properties.
+    
+    Override rubric takes precedence for descriptions and settings.
+    Weights are merged multiplicatively when both are non-default.
+    
+    Args:
+        base: Base rubric providing defaults
+        override: Override rubric with specific customizations
+        
+    Returns:
+        Blended rubric or None if both inputs are None
+    """
+    if override is None and base is None:
+        return None
+    if base is None:
+        return override
+    if override is None:
+        return base
+
+    base_map = {criterion.id: criterion for criterion in base.criteria}
+    merged: list[Criterion] = []
+
+    for ov in override.criteria:
+        if ov.id in base_map:
+            existing = base_map.pop(ov.id)
+            merged.append(
+                Criterion(
+                    id=ov.id,
+                    description=ov.description or existing.description,
+                    weight=_merge_weights(existing, ov),
+                    required=ov.required if ov.required is not None else existing.required,
+                )
+            )
+        else:
+            merged.append(ov)
+
+    merged.extend(base_map.values())
+
+    aggregation = override.aggregation
+    if aggregation == "inherit":
+        aggregation = base.aggregation
+
+    return Rubric(
+        version=override.version or base.version,
+        goal_text=override.goal_text or base.goal_text,
+        criteria=merged,
+        aggregation=aggregation,
+    )
+

synth_ai/task/rubrics/models.py

@@ -0,0 +1,57 @@
+"""Rubric and Criterion data models."""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field, field_validator
+
+
+class Criterion(BaseModel):
+    """Single scoring criterion within a rubric.
+    
+    Flexible variant allowing weights > 1.0 and no normalization requirement.
+    Used by task apps for general rubric scoring.
+    """
+
+    id: str
+    description: str
+    weight: float = 1.0
+    required: bool = False
+
+    @field_validator("weight")
+    @classmethod
+    def _validate_weight(cls, value: float) -> float:
+        if value <= 0:
+            raise ValueError("criterion weight must be positive")
+        return value
+
+
+class Rubric(BaseModel):
+    """Rubric definition for scoring task app outcomes.
+    
+    Supports flexible aggregation and blending. Criteria weights do not need
+    to sum to 1.0, making this suitable for general task app usage.
+    """
+
+    version: str
+    goal_text: str | None = None
+    criteria: list[Criterion] = Field(default_factory=list)
+    aggregation: str = "weighted_sum"
+
+    @field_validator("aggregation")
+    @classmethod
+    def _validate_aggregation(cls, value: str) -> str:
+        allowed = {"sum", "weighted_sum", "custom", "inherit"}
+        if value not in allowed:
+            raise ValueError(f"aggregation must be one of {sorted(allowed)}")
+        return value
+
+    @field_validator("criteria")
+    @classmethod
+    def _validate_criteria(cls, criteria: list[Criterion]) -> list[Criterion]:
+        seen = set()
+        for criterion in criteria:
+            if criterion.id in seen:
+                raise ValueError(f"duplicate criterion id: {criterion.id}")
+            seen.add(criterion.id)
+        return criteria
+

synth_ai/task/rubrics/scoring.py

@@ -0,0 +1,116 @@
+"""Rubric scoring utilities for events and outcomes."""
+
+from __future__ import annotations
+
+from collections.abc import Iterable
+from typing import Any
+
+from .models import Criterion, Rubric
+
+
+def _as_float(value: Any) -> float | None:
+    """Safely convert value to float, returning None on failure."""
+    try:
+        return float(value)
+    except Exception:
+        return None
+
+
+def _score(
+    criteria: Iterable[Criterion], values: dict[str, float], aggregation: str
+) -> dict[str, Any]:
+    """Compute aggregate score from criterion values.
+    
+    Args:
+        criteria: List of criteria defining scoring dimensions
+        values: Map of criterion IDs to scores
+        aggregation: How to aggregate ("sum", "weighted_sum", "custom")
+        
+    Returns:
+        Dict with aggregation method, total score, and per-criterion breakdown
+    """
+    if aggregation == "inherit":
+        aggregation = "weighted_sum"
+    per_criterion: dict[str, dict[str, Any]] = {}
+    total = 0.0
+    total_weight = 0.0
+    for criterion in criteria:
+        score = values.get(criterion.id, 0.0)
+        per_criterion[criterion.id] = {
+            "score": score,
+            "weight": criterion.weight,
+            "required": criterion.required,
+        }
+        if aggregation == "sum":
+            total += score
+        elif aggregation == "weighted_sum":
+            total += score * criterion.weight
+            total_weight += criterion.weight
+    if aggregation == "weighted_sum" and total_weight > 0:
+        total = total / total_weight
+    if aggregation == "custom":
+        total = None  # type: ignore[assignment]
+    return {
+        "aggregation": aggregation,
+        "score": total,
+        "per_criterion": per_criterion,
+    }
+
+
+def score_events_against_rubric(
+    events: list[dict[str, Any]], rubric: Rubric | None
+) -> dict[str, Any]:
+    """Score a list of evaluation events against a rubric.
+    
+    Events should contain criterion_id/id/criterion and score fields.
+    
+    Args:
+        events: List of event dicts with scoring info
+        rubric: Rubric defining criteria and aggregation
+        
+    Returns:
+        Scoring result with total and per-criterion scores
+    """
+    if rubric is None:
+        return {"aggregation": "none", "score": None, "per_criterion": {}}
+    values: dict[str, float] = {}
+    for event in events or []:
+        if not isinstance(event, dict):
+            continue
+        cid = event.get("criterion_id") or event.get("id") or event.get("criterion")
+        score = _as_float(event.get("score"))
+        if cid and score is not None:
+            values[str(cid)] = score
+    return _score(rubric.criteria, values, rubric.aggregation)
+
+
+def score_outcome_against_rubric(outcome: dict[str, Any], rubric: Rubric | None) -> dict[str, Any]:
+    """Score a rollout outcome against a rubric.
+    
+    Outcome should be a dict mapping criterion IDs to scores, optionally
+    nested under a "criteria" key.
+    
+    Args:
+        outcome: Outcome dict with criterion scores
+        rubric: Rubric defining criteria and aggregation
+        
+    Returns:
+        Scoring result with total and per-criterion scores
+    """
+    if rubric is None:
+        return {"aggregation": "none", "score": None, "per_criterion": {}}
+    values: dict[str, float] = {}
+    if isinstance(outcome, dict):
+        candidates = (
+            outcome.get("criteria") if isinstance(outcome.get("criteria"), dict) else outcome
+        )
+        if isinstance(candidates, dict):
+            for key, value in candidates.items():
+                score = _as_float(value)
+                if score is not None:
+                    values[str(key)] = score
+    return _score(rubric.criteria, values, rubric.aggregation)
+
+
+
+

synth_ai/{rubrics/validators.py → task/rubrics/strict.py}

@@ -1,15 +1,32 @@
+"""Strict rubric validators for step-wise judges.
+
+These validators enforce stricter constraints than the general-purpose rubrics:
+- Weights must be ≤ 1.0 and sum to exactly 1.0
+- Only weighted_sum aggregation is allowed
+- All required fields must be non-empty
+
+Used primarily for validation in judge configurations.
+"""
+
 from __future__ import annotations
 
 import json
 import math
+from collections.abc import Iterable
 from pathlib import Path
-from typing import Any, Iterable, Literal
+from typing import Any, Literal
 
 import pydantic
 
 
-class RubricCriterion(pydantic.BaseModel):
-    """Single scoring criterion within a rubric."""
+class StrictCriterion(pydantic.BaseModel):
+    """Single scoring criterion with strict validation.
+    
+    Enforces:
+    - Weight ≤ 1.0 (for proper normalization)
+    - Weight > 0.0 (positive)
+    - Non-empty strings
+    """
 
     id: str
     description: str
@@ -35,16 +52,23 @@ class RubricCriterion(pydantic.BaseModel):
         return value
 
 
-class RubricSpec(pydantic.BaseModel):
-    """High-level rubric definition used by step-wise judges."""
+class StrictRubric(pydantic.BaseModel):
+    """Strict rubric definition for step-wise judges.
+    
+    Enforces:
+    - Weights must sum to 1.0
+    - Only weighted_sum aggregation
+    - Non-empty version and goal_text
+    - At least one criterion
+    """
 
     version: str
     goal_text: str
     aggregation: Literal["weighted_sum"]
-    criteria: list[RubricCriterion]
+    criteria: list[StrictCriterion]
 
     @pydantic.model_validator(mode="after")
-    def _validate_weights(self) -> "RubricSpec":
+    def _validate_weights(self) -> StrictRubric:
         if not self.criteria:
             raise ValueError("rubric must declare at least one criterion")
         total_weight = sum(criterion.weight for criterion in self.criteria)
@@ -71,56 +95,55 @@ class RubricSpec(pydantic.BaseModel):
         return value
 
 
+# Re-export pydantic's ValidationError for convenience
 ValidationError = pydantic.ValidationError
 
 
-def validate_rubric_dict(payload: dict[str, Any]) -> RubricSpec:
-    """
-    Validate an in-memory rubric payload and return the parsed model.
-
+def validate_rubric_dict(payload: dict[str, Any]) -> StrictRubric:
+    """Validate an in-memory rubric payload with strict rules.
+    
     Args:
-        payload: Dictionary representing the rubric JSON.
+        payload: Dictionary representing the rubric JSON
+        
     Returns:
-        Validated RubricSpec instance.
+        Validated StrictRubric instance
+        
     Raises:
-        ValidationError: If the payload is missing required fields or contains
-        invalid weights.
+        ValidationError: If payload is invalid or doesn't meet strict constraints
     """
-
     if not isinstance(payload, dict):
         raise TypeError("rubric payload must be a dictionary")
-    return RubricSpec.model_validate(payload)
+    return StrictRubric.model_validate(payload)
 
 
 def _load_payload_from_file(path: Path) -> dict[str, Any]:
+    """Load JSON rubric from file."""
     if path.suffix.lower() != ".json":
         raise ValueError(f"Unsupported rubric file type: {path}")
     text = path.read_text(encoding="utf-8")
     return json.loads(text)
 
 
-def validate_rubric_file(path: Path) -> RubricSpec:
-    """
-    Load and validate a rubric file.
-
+def validate_rubric_file(path: Path) -> StrictRubric:
+    """Load and validate a rubric file with strict rules.
+    
     Args:
-        path: Path to a JSON rubric document.
+        path: Path to a JSON rubric document
+        
     Returns:
-        Validated RubricSpec instance.
+        Validated StrictRubric instance
     """
-
     payload = _load_payload_from_file(path)
     return validate_rubric_dict(payload)
 
 
-def validate_rubric_files(paths: Iterable[Path]) -> list[RubricSpec]:
-    """
-    Validate multiple rubric files and return their parsed models.
-
+def validate_rubric_files(paths: Iterable[Path]) -> list[StrictRubric]:
+    """Validate multiple rubric files with strict rules.
+    
     Useful for bulk validation inside tests or CI checks.
     """
-
-    validated: list[RubricSpec] = []
+    validated: list[StrictRubric] = []
     for path in paths:
         validated.append(validate_rubric_file(path))
     return validated
+

synth_ai/task/server.py

@@ -70,7 +70,7 @@ class TaskAppConfig:
     provide_task_instances: InstanceProvider
     rollout: RolloutExecutor
     dataset_registry: TaskDatasetRegistry | None = None
-    rubrics: RubricBundle = field(default_factory=RubricBundle)
+    rubrics: RubricBundle | None = field(default_factory=RubricBundle)
     proxy: ProxyConfig | None = None
     routers: Sequence[APIRouter] = field(default_factory=tuple)
     middleware: Sequence[Middleware] = field(default_factory=tuple)
@@ -93,7 +93,7 @@ class TaskAppConfig:
             provide_task_instances=self.provide_task_instances,
             rollout=self.rollout,
             dataset_registry=self.dataset_registry,
-            rubrics=self.rubrics,
+            rubrics=self.rubrics or RubricBundle(),
             proxy=self.proxy,
             routers=tuple(self.routers),
             middleware=tuple(self.middleware),
@@ -221,6 +221,7 @@ def _auth_dependency_factory(config: TaskAppConfig) -> Callable[[Request], None]
 
 def create_task_app(config: TaskAppConfig) -> FastAPI:
     cfg = config.clone()
+    cfg.rubrics = cfg.rubrics or RubricBundle()
     app = FastAPI(title=cfg.name, description=cfg.description)
 
     for key, value in cfg.app_state.items():
@@ -310,20 +311,20 @@ def create_task_app(config: TaskAppConfig) -> FastAPI:
     async def info() -> Mapping[str, Any]:
         dataset_meta = cfg.base_task_info.dataset
         rubrics: dict[str, Any] | None = None
-        if cfg.rubrics.outcome or cfg.rubrics.events:
+        rubric_bundle = cfg.rubrics
+        if rubric_bundle and (rubric_bundle.outcome or rubric_bundle.events):
             rubrics = {
-                "outcome": cfg.rubrics.outcome.model_dump() if cfg.rubrics.outcome else None,
-                "events": cfg.rubrics.events.model_dump() if cfg.rubrics.events else None,
+                "outcome": rubric_bundle.outcome.model_dump() if rubric_bundle.outcome else None,
+                "events": rubric_bundle.events.model_dump() if rubric_bundle.events else None,
             }
         payload = {
             "service": {
                 "task": cfg.base_task_info.task,
-                "version": cfg.base_task_info.task.get("version"),
+                "version": cfg.base_task_info.task.version,
             },
             "dataset": dataset_meta,
             "rubrics": rubrics,
             "inference": cfg.base_task_info.inference,
-            "capabilities": cfg.base_task_info.capabilities,
             "limits": cfg.base_task_info.limits,
         }
         return to_jsonable(payload)