synth-ai 0.2.9.dev11__py3-none-any.whl → 0.4.1__py3-none-any.whl

synth_ai/cli/commands/train/prompt_learning_validation.py

@@ -0,0 +1,633 @@
+"""Validation for prompt learning (GEPA/MIPRO) configurations.
+
+This module validates TOML configs and warns about:
+1. Unrecognized/unknown sections/fields that won't be processed
+2. Deprecated fields that should be migrated
+3. Missing required fields
+"""
+
+from __future__ import annotations
+
+import warnings
+from pathlib import Path
+from typing import Any
+
+# Known top-level TOML sections
+KNOWN_TOP_LEVEL_SECTIONS = {
+    "prompt_learning",
+    "display",  # Deprecated but recognized
+    "termination_config",  # Supported - creates TerminationManager conditions
+}
+
+# Known fields in [prompt_learning] section
+KNOWN_PROMPT_LEARNING_FIELDS = {
+    "algorithm",
+    "task_app_url",
+    "task_app_api_key",
+    "task_app_id",
+    "initial_prompt",
+    "policy",
+    "mipro",
+    "gepa",
+    "judge",
+    "proxy_models",
+    "env_config",
+    "env_name",
+    "termination_config",  # Supported - backend uses TerminationManager
+    "free_tier",  # Enable free tier mode with OSS models
+    # Deprecated fields (still recognized for warnings)
+    "results_folder",
+    "env_file_path",
+    # Seed pools (can be at top-level or nested)
+    "bootstrap_train_seeds",
+    "online_pool",
+    "test_pool",
+    "reference_pool",
+}
+
+# Known fields in [prompt_learning.policy] section
+KNOWN_POLICY_FIELDS = {
+    "model",
+    "provider",
+    "inference_url",
+    "inference_mode",
+    "temperature",
+    "max_completion_tokens",
+    "policy_name",
+}
+
+# Known fields in [prompt_learning.termination_config] section
+# All of these are SUPPORTED by backend's TerminationManager
+KNOWN_TERMINATION_CONFIG_FIELDS = {
+    "max_cost_usd",  # BudgetTerminationCondition
+    "max_trials",  # TrialLimitTerminationCondition
+    "max_seconds",  # TimeLimitTerminationCondition
+    "max_time_seconds",  # Alias for max_seconds
+    "max_rollouts",  # RolloutLimitTerminationCondition
+    "max_trials_without_improvement",  # NoImprovementTerminationCondition
+    "pessimism_enabled",  # PessimismTerminationCondition
+    "max_category_costs_usd",  # CategoryBudgetTerminationCondition
+}
+
+# Known fields in [prompt_learning.gepa] section
+KNOWN_GEPA_FIELDS = {
+    # Top-level GEPA fields
+    "env_name",
+    "env_config",
+    "rng_seed",
+    "proposer_type",
+    "proposer_effort",
+    "proposer_output_tokens",
+    "metaprompt",
+    "modules",
+    # Nested subsections
+    "rollout",
+    "evaluation",
+    "mutation",
+    "population",
+    "archive",
+    "token",
+    "judge",
+    "proxy_models",
+    "adaptive_pool",
+    "adaptive_batch",
+    # Backwards-compat flat fields (deprecated but recognized)
+    "rollout_budget",
+    "max_concurrent_rollouts",
+    "minibatch_size",
+    "evaluation_seeds",
+    "validation_seeds",
+    "test_pool",
+    "validation_pool",
+    "validation_top_k",
+    "mutation_rate",
+    "mutation_llm_model",
+    "mutation_llm_provider",
+    "mutation_llm_inference_url",
+    "mutation_prompt",
+    "initial_population_size",
+    "num_generations",
+    "children_per_generation",
+    "crossover_rate",
+    "selection_pressure",
+    "patience_generations",
+    "archive_size",
+    "pareto_set_size",
+    "pareto_eps",
+    "feedback_fraction",
+    "max_token_limit",
+    "token_counting_model",
+    "enforce_pattern_token_limit",
+    "max_spend_usd",
+}
+
+# Known fields in [prompt_learning.gepa.rollout]
+KNOWN_GEPA_ROLLOUT_FIELDS = {
+    "budget",
+    "max_concurrent",
+    "minibatch_size",
+}
+
+# Known fields in [prompt_learning.gepa.evaluation]
+KNOWN_GEPA_EVALUATION_FIELDS = {
+    "seeds",
+    "validation_seeds",
+    "test_pool",
+    "validation_pool",
+    "validation_top_k",
+}
+
+# Known fields in [prompt_learning.gepa.mutation]
+KNOWN_GEPA_MUTATION_FIELDS = {
+    "rate",
+    "llm_model",
+    "llm_provider",
+    "llm_inference_url",
+    "prompt",
+}
+
+# Known fields in [prompt_learning.gepa.population]
+KNOWN_GEPA_POPULATION_FIELDS = {
+    "initial_size",
+    "num_generations",
+    "children_per_generation",
+    "crossover_rate",
+    "selection_pressure",
+    "patience_generations",
+}
+
+# Known fields in [prompt_learning.gepa.archive]
+KNOWN_GEPA_ARCHIVE_FIELDS = {
+    "size",
+    "pareto_set_size",
+    "pareto_eps",
+    "feedback_fraction",
+}
+
+# Known fields in [prompt_learning.gepa.token]
+KNOWN_GEPA_TOKEN_FIELDS = {
+    "max_limit",
+    "counting_model",
+    "enforce_pattern_limit",
+    "max_spend_usd",
+}
+
+# Known fields in [prompt_learning.mipro] section
+KNOWN_MIPRO_FIELDS = {
+    "task_app_url",
+    "task_app_api_key",
+    "task_app_id",
+    "num_iterations",
+    "num_evaluations_per_iteration",
+    "batch_size",
+    "max_concurrent",
+    "env_name",
+    "env_config",
+    "meta_model",
+    "meta_model_provider",
+    "meta_model_inference_url",
+    "few_shot_score_threshold",
+    "results_file",
+    "max_wall_clock_seconds",
+    "max_total_tokens",
+    "policy_config",
+    "meta",
+    "modules",
+    "seeds",
+    "proposer_effort",
+    "proposer_output_tokens",
+    "max_token_limit",
+    "max_spend_usd",
+    "token_counting_model",
+    "enforce_token_limit",
+    "tpe",
+    "demo",
+    "grounding",
+    "meta_update",
+    "judge",
+    "proxy_models",
+    "adaptive_pool",
+    "spec_path",
+    "spec_max_tokens",
+    "spec_include_examples",
+    "spec_priority_threshold",
+    "metaprompt",
+    "bootstrap_train_seeds",
+    "online_pool",
+    "test_pool",
+    "reference_pool",
+    "min_bootstrap_demos",
+}
+
+# Known fields in [prompt_learning.judge]
+KNOWN_JUDGE_FIELDS = {
+    "enabled",
+    "reward_source",
+    "backend_base",
+    "backend_api_key_env",
+    "backend_provider",
+    "backend_model",
+    "synth_verifier_id",
+    "backend_rubric_id",
+    "backend_event_enabled",
+    "backend_outcome_enabled",
+    "backend_options",
+    "concurrency",
+    "timeout",
+    "weight_env",
+    "weight_event",
+    "weight_outcome",
+    "spec_path",
+    "spec_max_tokens",
+    "spec_context",
+}
+
+# Known fields in adaptive_pool config
+KNOWN_ADAPTIVE_POOL_FIELDS = {
+    "level",
+    "anchor_size",
+    "pool_init_size",
+    "pool_min_size",
+    "warmup_iters",
+    "anneal_stop_iter",
+    "pool_update_period",
+    "min_evals_per_example",
+    "k_info_prompts",
+    "info_buffer_factor",
+    "info_epsilon",
+    "anchor_selection_method",
+    "exploration_strategy",
+    "heatup_reserve_pool",
+    "heatup_trigger",
+    "heatup_size",
+    "heatup_cooldown_trials",
+    "heatup_schedule",
+}
+
+# Known fields in adaptive_batch config (GEPA only)
+KNOWN_ADAPTIVE_BATCH_FIELDS = {
+    "level",
+    "reflection_minibatch_size",
+    "min_local_improvement",
+    "val_evaluation_mode",
+    "val_subsample_size",
+    "candidate_selection_strategy",
+}
+
+# Known fields in proxy_models config
+KNOWN_PROXY_MODELS_FIELDS = {
+    "hi_provider",
+    "hi_model",
+    "lo_provider",
+    "lo_model",
+    "n_min_hi",
+    "r2_thresh",
+    "r2_stop",
+    "sigma_max",
+    "sigma_stop",
+    "verify_every",
+    "proxy_patience_usd",
+}
+
+# Deprecated fields with migration suggestions
+DEPRECATED_FIELDS = {
+    # Truly deprecated (backend ignores)
+    "display": "The [display] section is deprecated and ignored by the backend. Remove it from your config.",
+    "results_folder": "'results_folder' is deprecated and ignored by the backend. Remove it from your config.",
+    "env_file_path": "'env_file_path' is deprecated. Use environment variables instead.",
+    # Deprecated flat GEPA fields (prefer nested)
+    "rollout_budget": "Use [prompt_learning.gepa.rollout].budget instead of flat rollout_budget.",
+    "max_concurrent_rollouts": "Use [prompt_learning.gepa.rollout].max_concurrent instead.",
+    "evaluation_seeds": "Use [prompt_learning.gepa.evaluation].seeds instead of flat evaluation_seeds.",
+    "validation_seeds": "Use [prompt_learning.gepa.evaluation].validation_seeds instead.",
+    "backend_rubric_id": "Use 'synth_verifier_id' instead of 'backend_rubric_id' in [prompt_learning.judge].",
+}
+
+
+class PromptLearningValidationResult:
+    """Result of prompt learning config validation."""
+
+    def __init__(self) -> None:
+        self.errors: list[str] = []
+        self.warnings: list[str] = []
+        self.info: list[str] = []
+
+    @property
+    def is_valid(self) -> bool:
+        """Config is valid if there are no errors."""
+        return len(self.errors) == 0
+
+    def add_error(self, msg: str) -> None:
+        self.errors.append(msg)
+
+    def add_warning(self, msg: str) -> None:
+        self.warnings.append(msg)
+
+    def add_info(self, msg: str) -> None:
+        self.info.append(msg)
+
+    def __str__(self) -> str:
+        lines = []
+        if self.errors:
+            lines.append("ERRORS:")
+            for e in self.errors:
+                lines.append(f"  ❌ {e}")
+        if self.warnings:
+            lines.append("WARNINGS:")
+            for w in self.warnings:
+                lines.append(f"  ⚠️  {w}")
+        if self.info:
+            lines.append("INFO:")
+            for i in self.info:
+                lines.append(f"  ℹ️  {i}")
+        return "\n".join(lines) if lines else "✅ Config is valid"
+
+
+def _check_unknown_fields(
+    config: dict[str, Any],
+    known_fields: set[str],
+    section_path: str,
+    result: PromptLearningValidationResult,
+) -> None:
+    """Check for unknown fields in a config section."""
+    for key in config:
+        if key not in known_fields:
+            result.add_warning(
+                f"Unknown field '{key}' in [{section_path}]. "
+                f"This field will be ignored. Check spelling or remove it."
+            )
+
+
+def _check_deprecated_fields(
+    config: dict[str, Any],
+    section_path: str,
+    result: PromptLearningValidationResult,
+) -> None:
+    """Check for deprecated fields and suggest migrations."""
+    for key in config:
+        if key in DEPRECATED_FIELDS:
+            result.add_warning(f"[{section_path}] {DEPRECATED_FIELDS[key]}")
+
+
+def validate_prompt_learning_config(
+    config: dict[str, Any],
+    *,
+    config_path: Path | None = None,
+) -> PromptLearningValidationResult:
+    """Validate a prompt learning configuration.
+
+    Args:
+        config: Parsed TOML config dictionary
+        config_path: Optional path for better error messages
+
+    Returns:
+        PromptLearningValidationResult with errors, warnings, and info
+    """
+    result = PromptLearningValidationResult()
+    path_prefix = f"({config_path}) " if config_path else ""
+
+    # Check top-level sections
+    for key in config:
+        if key not in KNOWN_TOP_LEVEL_SECTIONS:
+            result.add_warning(
+                f"{path_prefix}Unknown top-level section '[{key}]'. "
+                f"Known sections: {', '.join(sorted(KNOWN_TOP_LEVEL_SECTIONS))}"
+            )
+
+    # Check for deprecated [display] section
+    if "display" in config:
+        result.add_warning(
+            f"{path_prefix}The [display] section is deprecated and ignored by the backend. "
+            "Remove it to clean up your config."
+        )
+
+    # Validate [prompt_learning] section
+    pl_config = config.get("prompt_learning", {})
+    if not pl_config:
+        result.add_error(f"{path_prefix}Missing required [prompt_learning] section")
+        return result
+
+    # Check for unknown fields in prompt_learning
+    _check_unknown_fields(pl_config, KNOWN_PROMPT_LEARNING_FIELDS, "prompt_learning", result)
+
+    # Check for deprecated fields
+    _check_deprecated_fields(pl_config, "prompt_learning", result)
+
+    # Check required fields
+    algorithm = pl_config.get("algorithm")
+    if not algorithm:
+        result.add_error(f"{path_prefix}Missing required 'algorithm' field in [prompt_learning]")
+    elif algorithm not in ("gepa", "mipro"):
+        result.add_error(
+            f"{path_prefix}Invalid algorithm '{algorithm}'. Must be 'gepa' or 'mipro'"
+        )
+
+    if not pl_config.get("task_app_url"):
+        result.add_error(f"{path_prefix}Missing required 'task_app_url' in [prompt_learning]")
+
+    # Validate [prompt_learning.policy] if present
+    policy = pl_config.get("policy")
+    if policy and isinstance(policy, dict):
+        _check_unknown_fields(policy, KNOWN_POLICY_FIELDS, "prompt_learning.policy", result)
+
+    # Validate [prompt_learning.termination_config] if present
+    termination = pl_config.get("termination_config")
+    if termination and isinstance(termination, dict):
+        _check_unknown_fields(
+            termination,
+            KNOWN_TERMINATION_CONFIG_FIELDS,
+            "prompt_learning.termination_config",
+            result,
+        )
+        # Info: termination_config IS supported
+        result.add_info(
+            "termination_config is supported and will create backend TerminationManager conditions"
+        )
+
+    # Validate [prompt_learning.judge] if present
+    judge = pl_config.get("judge")
+    if judge and isinstance(judge, dict):
+        _check_unknown_fields(judge, KNOWN_JUDGE_FIELDS, "prompt_learning.judge", result)
+
+    # Validate [prompt_learning.proxy_models] if present
+    proxy_models = pl_config.get("proxy_models")
+    if proxy_models and isinstance(proxy_models, dict):
+        _check_unknown_fields(
+            proxy_models, KNOWN_PROXY_MODELS_FIELDS, "prompt_learning.proxy_models", result
+        )
+
+    # Validate algorithm-specific sections
+    if algorithm == "gepa":
+        _validate_gepa_config(pl_config.get("gepa", {}), result, path_prefix)
+    elif algorithm == "mipro":
+        _validate_mipro_config(pl_config.get("mipro", {}), result, path_prefix)
+
+    return result
+
+
+def _validate_gepa_config(
+    gepa: dict[str, Any],
+    result: PromptLearningValidationResult,
+    path_prefix: str,
+) -> None:
+    """Validate GEPA-specific configuration."""
+    if not gepa:
+        result.add_warning(f"{path_prefix}No [prompt_learning.gepa] section found for GEPA algorithm")
+        return
+
+    # Check for unknown fields
+    _check_unknown_fields(gepa, KNOWN_GEPA_FIELDS, "prompt_learning.gepa", result)
+
+    # Check for deprecated flat fields
+    deprecated_flat = {
+        "rollout_budget",
+        "max_concurrent_rollouts",
+        "evaluation_seeds",
+        "validation_seeds",
+    }
+    for field in deprecated_flat:
+        if field in gepa:
+            result.add_info(
+                f"Using flat '{field}' in [prompt_learning.gepa] - "
+                "consider migrating to nested structure for clarity"
+            )
+
+    # Validate nested sections
+    if "rollout" in gepa and isinstance(gepa["rollout"], dict):
+        _check_unknown_fields(
+            gepa["rollout"], KNOWN_GEPA_ROLLOUT_FIELDS, "prompt_learning.gepa.rollout", result
+        )
+
+    if "evaluation" in gepa and isinstance(gepa["evaluation"], dict):
+        _check_unknown_fields(
+            gepa["evaluation"],
+            KNOWN_GEPA_EVALUATION_FIELDS,
+            "prompt_learning.gepa.evaluation",
+            result,
+        )
+
+    if "mutation" in gepa and isinstance(gepa["mutation"], dict):
+        _check_unknown_fields(
+            gepa["mutation"], KNOWN_GEPA_MUTATION_FIELDS, "prompt_learning.gepa.mutation", result
+        )
+
+    if "population" in gepa and isinstance(gepa["population"], dict):
+        _check_unknown_fields(
+            gepa["population"],
+            KNOWN_GEPA_POPULATION_FIELDS,
+            "prompt_learning.gepa.population",
+            result,
+        )
+
+    if "archive" in gepa and isinstance(gepa["archive"], dict):
+        _check_unknown_fields(
+            gepa["archive"], KNOWN_GEPA_ARCHIVE_FIELDS, "prompt_learning.gepa.archive", result
+        )
+
+    if "token" in gepa and isinstance(gepa["token"], dict):
+        _check_unknown_fields(
+            gepa["token"], KNOWN_GEPA_TOKEN_FIELDS, "prompt_learning.gepa.token", result
+        )
+
+    if "adaptive_pool" in gepa and isinstance(gepa["adaptive_pool"], dict):
+        _check_unknown_fields(
+            gepa["adaptive_pool"],
+            KNOWN_ADAPTIVE_POOL_FIELDS,
+            "prompt_learning.gepa.adaptive_pool",
+            result,
+        )
+
+    if "adaptive_batch" in gepa and isinstance(gepa["adaptive_batch"], dict):
+        _check_unknown_fields(
+            gepa["adaptive_batch"],
+            KNOWN_ADAPTIVE_BATCH_FIELDS,
+            "prompt_learning.gepa.adaptive_batch",
+            result,
+        )
+
+    if "proxy_models" in gepa and isinstance(gepa["proxy_models"], dict):
+        _check_unknown_fields(
+            gepa["proxy_models"],
+            KNOWN_PROXY_MODELS_FIELDS,
+            "prompt_learning.gepa.proxy_models",
+            result,
+        )
+
+    if "judge" in gepa and isinstance(gepa["judge"], dict):
+        _check_unknown_fields(
+            gepa["judge"], KNOWN_JUDGE_FIELDS, "prompt_learning.gepa.judge", result
+        )
+
+
+def _validate_mipro_config(
+    mipro: dict[str, Any],
+    result: PromptLearningValidationResult,
+    path_prefix: str,
+) -> None:
+    """Validate MIPRO-specific configuration."""
+    if not mipro:
+        result.add_warning(
+            f"{path_prefix}No [prompt_learning.mipro] section found for MIPRO algorithm"
+        )
+        return
+
+    # Check for unknown fields
+    _check_unknown_fields(mipro, KNOWN_MIPRO_FIELDS, "prompt_learning.mipro", result)
+
+    # Validate nested sections
+    if "judge" in mipro and isinstance(mipro["judge"], dict):
+        _check_unknown_fields(
+            mipro["judge"], KNOWN_JUDGE_FIELDS, "prompt_learning.mipro.judge", result
+        )
+
+    if "adaptive_pool" in mipro and isinstance(mipro["adaptive_pool"], dict):
+        _check_unknown_fields(
+            mipro["adaptive_pool"],
+            KNOWN_ADAPTIVE_POOL_FIELDS,
+            "prompt_learning.mipro.adaptive_pool",
+            result,
+        )
+
+    if "proxy_models" in mipro and isinstance(mipro["proxy_models"], dict):
+        _check_unknown_fields(
+            mipro["proxy_models"],
+            KNOWN_PROXY_MODELS_FIELDS,
+            "prompt_learning.mipro.proxy_models",
+            result,
+        )
+
+
+def validate_and_warn(config: dict[str, Any], config_path: Path | None = None) -> None:
+    """Validate config and print warnings/errors to stderr.
+
+    This is a convenience function for CLI usage that prints validation
+    results and raises an exception if there are errors.
+
+    Args:
+        config: Parsed TOML config dictionary
+        config_path: Optional path for better error messages
+
+    Raises:
+        ValueError: If config has validation errors
+    """
+    result = validate_prompt_learning_config(config, config_path=config_path)
+
+    # Print warnings (these don't stop execution)
+    for warning in result.warnings:
+        warnings.warn(warning, UserWarning, stacklevel=2)
+
+    # Print info messages
+    for info in result.info:
+        print(f"ℹ️  {info}")
+
+    # Raise on errors
+    if result.errors:
+        error_msg = "\n".join(f"  - {e}" for e in result.errors)
+        raise ValueError(f"Config validation failed:\n{error_msg}")
+
+
+__all__ = [
+    "validate_prompt_learning_config",
+    "validate_and_warn",
+    "PromptLearningValidationResult",
+]