synth-ai 0.2.14__py3-none-any.whl → 0.4.1__py3-none-any.whl

synth_ai/products/graph_evolve/config.py

@@ -0,0 +1,591 @@
+"""Configuration for Graph Optimization jobs.
+
+This module provides Pydantic models for loading and validating
+graph optimization configuration from TOML files.
+
+Supports multiple algorithms:
+- graph_gepa: Grammatical Evolution for graph structure optimization
+"""
+
+from __future__ import annotations
+
+from enum import Enum
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+try:
+    import tomllib
+except ImportError:
+    import tomli as tomllib  # type: ignore[import-not-found]
+
+from pydantic import BaseModel, Field, field_validator
+
+
+class GraphType(str, Enum):
+    """What the graph does.
+
+    - POLICY: Maps inputs to outputs (standard LLM pipeline)
+    - VERIFIER: Judges/scores existing results
+    - RLM: Recursive Language Model - handles massive context (1M+ tokens) by keeping
+           it out of prompts and searching via tools. Auto-adds materialize_context,
+           local_grep, local_search, query_lm, and codex_exec tools.
+    """
+    POLICY = "policy"      # Maps inputs to outputs, solves tasks
+    VERIFIER = "verifier"  # Judges/scores existing results
+    RLM = "rlm"            # Recursive LM - massive context via tool search
+
+
+class GraphPattern(str, Enum):
+    """Architectural patterns for graph structure.
+
+    These patterns can be applied to ANY graph type (policy, verifier, rlm).
+    They describe HOW the graph processes data, not WHAT it does.
+
+    Use patterns to guide the proposer toward specific architectures:
+    - patterns.required=["rlm"] → MUST use RLM pattern (auto-adds tools)
+    - patterns.optional=["rlm", "map_reduce"] → May try either pattern
+    - patterns.prefer=["map_reduce"] → Prefer this pattern when viable
+
+    Patterns:
+    - RLM: Tool-based search for massive context (1M+ tokens).
+           Materializes context to files, uses grep/search tools.
+           Good for: RAG, codebase search, document QA.
+
+    - MAP_REDUCE: Parallel processing for variable-length inputs.
+           Maps over items in parallel, then reduces/aggregates.
+           Good for: scoring events, processing lists, chunking.
+
+    - SINGLE_SHOT: Single LLM call, minimal structure.
+           Good for: classification, simple QA.
+
+    - CHAIN_OF_THOUGHT: Multi-step reasoning, sequential nodes.
+           Good for: complex reasoning, multi-hop QA.
+
+    - DIGEST_COMBINE: Two-stage: digest in parallel, then combine.
+           Good for: verifiers analyzing multiple aspects.
+    """
+    RLM = "rlm"
+    MAP_REDUCE = "map_reduce"
+    SINGLE_SHOT = "single_shot"
+    CHAIN_OF_THOUGHT = "chain_of_thought"
+    DIGEST_COMBINE = "digest_combine"
+
+
+class PatternConfig(BaseModel):
+    """Configuration for which architectural patterns the proposer should use.
+
+    Patterns are orthogonal to graph_type - you can have an RLM-pattern verifier
+    or a map-reduce-pattern policy. This lets you guide the proposer toward
+    specific architectures without changing the fundamental graph type.
+
+    Examples:
+        # RLM verifier (MUST use RLM pattern - auto-adds tools & guidance)
+        patterns = PatternConfig(required=["rlm"])
+
+        # Let proposer try different patterns
+        patterns = PatternConfig(optional=["rlm", "map_reduce", "digest_combine"])
+
+        # Prefer map-reduce but allow alternatives
+        patterns = PatternConfig(
+            required=[],
+            optional=["map_reduce", "digest_combine"],
+            prefer=["map_reduce"]
+        )
+
+    When RLM pattern is in required/prefer:
+        - Tools are auto-added: materialize_context, local_grep, local_search, etc.
+        - Proposer receives RLM-specific guidance for tool-based search patterns
+    """
+    required: List[str] = Field(
+        default_factory=list,
+        description="Patterns the graph MUST use. Proposer will incorporate all required patterns."
+    )
+    optional: List[str] = Field(
+        default_factory=list,
+        description="Patterns the proposer MAY consider. These are suggestions, not requirements."
+    )
+    prefer: List[str] = Field(
+        default_factory=list,
+        description="Patterns to prefer when multiple options are viable."
+    )
+
+    def to_api_dict(self) -> Dict[str, Any]:
+        """Convert to API request format."""
+        return {
+            "required": [GraphPattern(p).value for p in self.required],
+            "optional": [GraphPattern(p).value for p in self.optional],
+            "prefer": [GraphPattern(p).value for p in self.prefer],
+        }
+
+
+class GraphStructure(str, Enum):
+    """Structural complexity of the graph."""
+    SINGLE_PROMPT = "single_prompt"  # One LLM call, minimal structure
+    DAG = "dag"                      # Multiple nodes in sequence, no branching
+    CONDITIONAL = "conditional"      # Full graph with conditional branching
+
+
+class ProposerConfig(BaseModel):
+    """Configuration for the LLM proposer."""
+    model: str = Field(default="gpt-4.1", description="Model for proposing patches")
+    temperature: float = Field(default=0.7, ge=0.0, le=2.0)
+    max_tokens: int = Field(default=4096, gt=0)
+
+
+class EvolutionConfig(BaseModel):
+    """Evolution algorithm parameters."""
+    num_generations: int = Field(default=5, ge=1, description="Number of evolution generations")
+    children_per_generation: int = Field(default=3, ge=1, description="Children per generation")
+
+
+class SeedsConfig(BaseModel):
+    """Train and validation seed configuration."""
+    train: List[int] = Field(default_factory=lambda: list(range(10)))
+    validation: List[int] = Field(default_factory=lambda: list(range(100, 105)))
+
+
+class LimitsConfig(BaseModel):
+    """Resource limits for the job."""
+    max_spend_usd: float = Field(default=10.0, gt=0)
+    timeout_seconds: int = Field(default=3600, gt=0)
+
+
+class IndifferencePointConfig(BaseModel):
+    """Defines trade-off equivalences at a specific anchor point.
+    
+    Example: At 80% accuracy, 2s latency, $0.50 cost:
+      - +2% accuracy change is equivalent to -0.4s latency or -$0.10 cost
+      - Differences below 0.5% accuracy are considered noise
+    """
+    # Anchor point (where trade-offs are defined)
+    reward: float = Field(..., description="Anchor reward (e.g., 0.80 for 80%)")
+    latency_s: float = Field(..., description="Anchor latency in seconds")
+    cost_usd: float = Field(..., description="Anchor cost per seed in USD")
+    
+    # Trade-off equivalences at this anchor
+    reward_delta: float = Field(default=0.02, description="Reward change considered equivalent")
+    latency_delta: float = Field(default=0.4, description="Latency change (s) equivalent to reward_delta")
+    cost_delta: float = Field(default=0.10, description="Cost change ($) equivalent to reward_delta")
+    
+    # Noise floors for each objective (differences smaller than these are ignored)
+    reward_noise: float = Field(default=0.005, description="Reward diffs below this are noise (e.g., 0.5%)")
+    latency_noise: float = Field(default=0.1, description="Latency diffs below this (s) are noise")
+    cost_noise: float = Field(default=0.01, description="Cost diffs below this ($) are noise")
+
+
+class ParetoFloorsConfig(BaseModel):
+    """Configuration for multi-objective Pareto comparison.
+
+    Controls:
+    1. Which objectives are included in Pareto comparison (enable flags)
+    2. Soft floors below which differences are ignored (indifference regions)
+    3. Hard ceilings that disqualify candidates entirely (budget constraints)
+    """
+    # Enable/disable objectives in Pareto comparison
+    use_latency: bool = Field(default=True, description="Include latency in Pareto comparison")
+    use_cost: bool = Field(default=True, description="Include cost in Pareto comparison")
+
+    # Soft floors: below these, all values are "equally good"
+    latency_s: float = Field(default=2.0, description="Don't discriminate on latency below this (s)")
+    cost_usd: float = Field(default=0.10, description="Don't discriminate on cost below this ($/seed)")
+
+    # Hard ceilings: disqualify candidates exceeding these limits
+    max_latency_s: Optional[float] = Field(default=None, description="Disqualify if mean latency > this")
+    max_cost_usd: Optional[float] = Field(default=None, description="Disqualify if mean cost/seed > this")
+    min_reward: Optional[float] = Field(default=None, description="Disqualify if mean reward < this")
+
+
+# ============================================================================
+# ADAS Dataset Format Models
+# ============================================================================
+
+class TaskInput(BaseModel):
+    """A single task/example in an ADAS dataset.
+
+    For POLICY graphs: Contains the problem to solve.
+    For VERIFIER graphs: Contains a trace to evaluate.
+
+    Example (Policy - QA):
+        {
+            "task_id": "q123",
+            "input": {
+                "question": "What is the capital of France?",
+                "context": "Paris is the capital and largest city of France."
+            }
+        }
+
+    Example (Verifier - Game traces):
+        {
+            "task_id": "trace_001",
+            "input": {
+                "game_state": {...},
+                "agent_action": "move_north",
+                "outcome": "success"
+            }
+        }
+    """
+    task_id: Optional[str] = Field(default=None, description="Unique identifier for this task")
+    id: Optional[str] = Field(default=None, description="Alternate ID field (task_id preferred)")
+    input: Dict[str, Any] = Field(..., description="Input data passed to the graph")
+
+    @field_validator("input", mode="before")
+    @classmethod
+    def ensure_input_dict(cls, v: Any) -> Dict[str, Any]:
+        if v is None:
+            return {}
+        if not isinstance(v, dict):
+            raise ValueError(f"Task input must be a dict, got {type(v).__name__}")
+        return v
+
+    def get_task_id(self) -> str:
+        """Get the task ID, preferring task_id over id."""
+        return self.task_id or self.id or "unknown"
+
+
+class GoldOutput(BaseModel):
+    """Ground truth for scoring a task.
+
+    For POLICY graphs: The expected answer/output.
+    For VERIFIER graphs: The gold score (0.0-1.0) from human evaluation.
+
+    Example (Policy - QA):
+        {
+            "task_id": "q123",
+            "output": {"answer": "Paris"},
+            "score": 1.0  # Optional: gold score for the expected output
+        }
+
+    Example (Verifier calibration):
+        {
+            "task_id": "trace_001",
+            "output": {},  # May be empty for verifiers
+            "score": 0.75  # Human calibration score
+        }
+    """
+    task_id: Optional[str] = Field(default=None, description="Must match a TaskInput.task_id")
+    output: Dict[str, Any] = Field(default_factory=dict, description="Expected output fields")
+    score: Optional[float] = Field(default=None, ge=0.0, le=1.0, description="Gold score (0.0-1.0)")
+
+
+class ADASDatasetMetadata(BaseModel):
+    """Metadata about an ADAS dataset.
+
+    Provides context for graph generation and optimization.
+    """
+    name: Optional[str] = Field(default=None, description="Dataset name/identifier")
+    task_description: Optional[str] = Field(default=None, description="What task this dataset represents")
+    input_schema: Optional[Dict[str, Any]] = Field(default=None, description="Schema of task inputs")
+    output_schema: Optional[Dict[str, Any]] = Field(default=None, description="Schema of expected outputs")
+    domain: Optional[str] = Field(default=None, description="Domain (qa, code, games, etc.)")
+
+
+class ADASDataset(BaseModel):
+    """Complete ADAS dataset format for inline upload.
+
+    This is the schema for the `dataset` field in GraphOptimizationConfig
+    when uploading data directly instead of using a pre-registered dataset.
+
+    Example:
+        {
+            "tasks": [
+                {"task_id": "q1", "input": {"question": "What is 2+2?"}},
+                {"task_id": "q2", "input": {"question": "What is 3+3?"}}
+            ],
+            "gold_outputs": [
+                {"task_id": "q1", "output": {"answer": "4"}, "score": 1.0},
+                {"task_id": "q2", "output": {"answer": "6"}, "score": 1.0}
+            ],
+            "metadata": {
+                "name": "simple_math",
+                "task_description": "Answer basic math questions"
+            }
+        }
+    """
+    tasks: List[TaskInput] = Field(..., min_length=1, description="List of tasks/examples")
+    gold_outputs: List[GoldOutput] = Field(..., min_length=1, description="Ground truth for each task")
+    metadata: ADASDatasetMetadata = Field(default_factory=ADASDatasetMetadata)
+
+    @field_validator("tasks", mode="before")
+    @classmethod
+    def validate_tasks(cls, v: Any) -> List[Dict[str, Any]]:
+        if not isinstance(v, list):
+            raise ValueError(f"tasks must be a list, got {type(v).__name__}")
+        if len(v) == 0:
+            raise ValueError("tasks list cannot be empty")
+        return v
+
+    @field_validator("gold_outputs", mode="before")
+    @classmethod
+    def validate_gold_outputs(cls, v: Any) -> List[Dict[str, Any]]:
+        if not isinstance(v, list):
+            raise ValueError(f"gold_outputs must be a list, got {type(v).__name__}")
+        if len(v) == 0:
+            raise ValueError("gold_outputs list cannot be empty")
+        return v
+
+    def validate_task_ids(self) -> List[str]:
+        """Validate that gold_outputs reference valid task IDs.
+
+        Returns list of warnings (non-fatal issues).
+        """
+        warnings = []
+        task_ids = {t.get_task_id() for t in self.tasks}
+        gold_task_ids = {g.task_id for g in self.gold_outputs if g.task_id}
+
+        # Check for gold outputs without matching tasks
+        orphan_golds = gold_task_ids - task_ids
+        if orphan_golds:
+            warnings.append(
+                f"Gold outputs reference unknown task IDs: {list(orphan_golds)[:5]}"
+            )
+
+        # Check for tasks without gold outputs
+        missing_golds = task_ids - gold_task_ids
+        if missing_golds:
+            warnings.append(
+                f"Tasks without gold outputs: {list(missing_golds)[:5]}"
+            )
+
+        return warnings
+
+
+class GraphOptimizationConfig(BaseModel):
+    """Complete configuration for a graph optimization job.
+    
+    Example TOML:
+        [graph_optimization]
+        algorithm = "graph_gepa"
+        dataset_name = "hotpotqa"
+        graph_type = "policy"
+        graph_structure = "dag"
+        
+        [graph_optimization.evolution]
+        num_generations = 5
+        children_per_generation = 3
+        
+        [graph_optimization.proposer]
+        model = "gpt-4.1"
+        
+        [graph_optimization.seeds]
+        train = [0, 1, 2, 3, 4]
+        validation = [100, 101, 102]
+        
+        [graph_optimization.limits]
+        max_spend_usd = 10.0
+    """
+    
+    # Algorithm selection
+    algorithm: str = Field(default="graph_gepa", description="Optimization algorithm (currently: 'graph_gepa')")
+    
+    # Required
+    dataset_name: str = Field(..., description="Dataset to optimize for (e.g., 'hotpotqa')")
+    
+    # Graph configuration
+    graph_type: GraphType = Field(default=GraphType.POLICY)
+    graph_structure: GraphStructure = Field(default=GraphStructure.DAG)
+    
+    # Custom topology guidance (adds detail to graph_structure, doesn't replace it)
+    topology_guidance: Optional[str] = Field(
+        default=None,
+        description="Additional guidance on what kind of graph to build within the chosen structure (e.g., 'Use a single LLM call that reasons and answers in one shot')"
+    )
+
+    # Pattern configuration - architectural patterns orthogonal to graph_type
+    patterns: Optional[PatternConfig] = Field(
+        default=None,
+        description=(
+            "Configure which architectural patterns the proposer should use/consider. "
+            "Patterns are orthogonal to graph_type - you can have an RLM-pattern verifier. "
+            "Example: patterns=PatternConfig(required=['rlm']) for RLM-style verifier."
+        )
+    )
+
+    # Optional warm start from a saved graph in the registry
+    initial_graph_id: Optional[str] = Field(
+        default=None,
+        description="Optional graph_id from the graphs registry to warm-start evolution.",
+    )
+
+    # Allowed policy models - which models the generated graph can use
+    allowed_policy_models: List[str] = Field(
+        default_factory=lambda: ["gpt-4o-mini", "gpt-4o"],
+        description="Models the graph is allowed to use in its nodes"
+    )
+    
+    # Nested configs
+    evolution: EvolutionConfig = Field(default_factory=EvolutionConfig)
+    proposer: ProposerConfig = Field(default_factory=ProposerConfig)
+    seeds: SeedsConfig = Field(default_factory=SeedsConfig)
+    limits: LimitsConfig = Field(default_factory=LimitsConfig)
+    
+    # Multi-objective Pareto configuration
+    indifference_points: List[IndifferencePointConfig] = Field(
+        default_factory=list,
+        description="Trade-off equivalences at anchor points"
+    )
+    pareto_floors: Optional[ParetoFloorsConfig] = Field(
+        default=None,
+        description="Thresholds below which metric differences are ignored"
+    )
+    
+    # Optional dataset-specific config
+    dataset_config: Dict[str, Any] = Field(default_factory=dict)
+
+    # Constraint: max LLM calls per execution
+    max_llm_calls_per_run: Optional[int] = Field(
+        default=None,
+        ge=1,
+        description="Maximum LLM calls allowed per graph execution (e.g., 1, 2, 5).",
+    )
+
+    
+    # Inline dataset upload (for verifier calibration, custom datasets)
+    # Format: {"name": str, "task_description": str, "examples": [...]}
+    dataset: Optional[Dict[str, Any]] = Field(
+        default=None,
+        description="Inline dataset for upload (ADAS format). If provided, dataset_name is used as identifier."
+    )
+    
+    # Task context for initial graph generation (when dataset doesn't provide it)
+    task_description: Optional[str] = Field(default=None, description="Description of the task")
+    input_schema: Optional[Dict[str, Any]] = Field(default=None, description="Expected input format")
+    output_schema: Optional[Dict[str, Any]] = Field(default=None, description="Expected output format")
+
+    # Problem specification - detailed task info for the graph proposer
+    # This should include domain-specific constraints (e.g., valid labels for classification tasks)
+    problem_spec: Optional[str] = Field(
+        default=None,
+        description=(
+            "Detailed problem specification for the graph proposer. "
+            "Include domain-specific information like valid output labels, constraints, "
+            "and any other information needed to generate correct graphs. "
+            "If provided, this is combined with task_description for the proposer context."
+        )
+    )
+    
+    # Scoring configuration
+    scoring_strategy: str = Field(default="rubric", description="Scoring strategy: 'default', 'rubric', 'mae'")
+    judge_model: str = Field(default="gpt-4o-mini", description="Model for LLM judge scoring")
+    
+    @field_validator("graph_type", mode="before")
+    @classmethod
+    def validate_graph_type(cls, v: Any) -> GraphType:
+        if isinstance(v, str):
+            return GraphType(v.lower())
+        return v
+    
+    @field_validator("graph_structure", mode="before")
+    @classmethod
+    def validate_graph_structure(cls, v: Any) -> GraphStructure:
+        if isinstance(v, str):
+            return GraphStructure(v.lower())
+        return v
+    
+    @classmethod
+    def from_toml(cls, path: str | Path) -> "GraphOptimizationConfig":
+        """Load configuration from a TOML file.
+        
+        Args:
+            path: Path to the TOML configuration file
+            
+        Returns:
+            Parsed GraphOptimizationConfig
+            
+        Raises:
+            FileNotFoundError: If config file doesn't exist
+            ValueError: If config is invalid
+        """
+        path = Path(path)
+        if not path.exists():
+            raise FileNotFoundError(f"Config file not found: {path}")
+        
+        with open(path, "rb") as f:
+            data = tomllib.load(f)
+        
+        # Extract graph_optimization section
+        if "graph_optimization" not in data:
+            raise ValueError(
+                f"Config file must have a [graph_optimization] section. "
+                f"Found sections: {list(data.keys())}"
+            )
+        
+        config_data = data["graph_optimization"]
+        return cls(**config_data)
+    
+    def to_request_dict(self) -> Dict[str, Any]:
+        """Convert config to API request format.
+        
+        Returns:
+            Dictionary suitable for POST to /graph-gepa/jobs
+        """
+        request = {
+            "dataset_name": self.dataset_name,
+            "train_seeds": self.seeds.train,
+            "val_seeds": self.seeds.validation,
+            "num_generations": self.evolution.num_generations,
+            "children_per_generation": self.evolution.children_per_generation,
+            "proposer_model": self.proposer.model,
+            "graph_type": self.graph_type.value,
+            "graph_structure": self.graph_structure.value,
+            "allowed_policy_models": self.allowed_policy_models,
+            "dataset_config": self.dataset_config,
+            "scoring_strategy": self.scoring_strategy,
+            "judge_model": self.judge_model,
+        }
+
+        if self.max_llm_calls_per_run is not None:
+            request["max_llm_calls_per_run"] = int(self.max_llm_calls_per_run)
+        
+        # Only include topology_guidance if set
+        if self.topology_guidance:
+            request["topology_guidance"] = self.topology_guidance
+
+        # Include pattern configuration if set
+        if self.patterns:
+            request["patterns"] = self.patterns.to_api_dict()
+
+        if self.initial_graph_id:
+            request["initial_graph_id"] = self.initial_graph_id
+        
+        # Inline dataset upload (for verifier calibration, custom datasets)
+        if self.dataset:
+            # Validate dataset structure using Pydantic model
+            try:
+                validated = ADASDataset(**self.dataset)
+                # Check for task ID consistency (non-fatal warnings)
+                warnings = validated.validate_task_ids()
+                if warnings:
+                    import logging
+                    logger = logging.getLogger(__name__)
+                    for w in warnings:
+                        logger.warning(f"[ADASDataset] {w}")
+            except Exception as e:
+                raise ValueError(
+                    f"Invalid ADAS dataset format: {e}\n"
+                    f"Expected format: {{'tasks': [...], 'gold_outputs': [...], 'metadata': {{...}}}}\n"
+                    f"See ADASDataset model for full schema.\n"
+                    f"Got keys: {list(self.dataset.keys())}"
+                )
+            request["dataset"] = self.dataset
+        
+        # Task context for initial graph generation
+        if self.task_description:
+            request["task_description"] = self.task_description
+        if self.input_schema:
+            request["input_schema"] = self.input_schema
+        if self.output_schema:
+            request["output_schema"] = self.output_schema
+        if self.problem_spec:
+            request["problem_spec"] = self.problem_spec
+
+        # Include indifference points for epsilon-Pareto dominance
+        if self.indifference_points:
+            request["indifference_points"] = [
+                p.model_dump() for p in self.indifference_points
+            ]
+        
+        # Include pareto floors for noise reduction
+        if self.pareto_floors:
+            request["pareto_floors"] = self.pareto_floors.model_dump()
+        
+        return request

synth_ai/products/graph_evolve/converters/__init__.py

@@ -0,0 +1,42 @@
+"""Dataset converters for Graph GEPA.
+
+This module provides converters to transform common dataset formats
+into ADAS format for use with Graph GEPA optimization.
+
+Supported formats:
+- OpenAI SFT: JSONL with messages array (system, user, assistant roles)
+
+Example:
+    >>> from synth_ai.products.graph_gepa.converters import convert_openai_sft
+    >>>
+    >>> # Convert from file
+    >>> result = convert_openai_sft("training_data.jsonl")
+    >>> adas_dataset = result.dataset
+    >>>
+    >>> # Use in GraphOptimizationConfig
+    >>> from synth_ai.products.graph_gepa import GraphOptimizationConfig
+    >>> config = GraphOptimizationConfig(
+    ...     dataset_name="my_qa_task",
+    ...     dataset=adas_dataset,
+    ...     graph_type="policy",
+    ...     ...
+    ... )
+"""
+
+from __future__ import annotations
+
+from .openai_sft import (
+    ConversionError,
+    ConversionResult,
+    ConversionWarning,
+    convert_openai_sft,
+    preview_conversion,
+)
+
+__all__ = [
+    "convert_openai_sft",
+    "preview_conversion",
+    "ConversionResult",
+    "ConversionWarning",
+    "ConversionError",
+]