"""ACELiteLLM — batteries-included conversational agent with ACE learning.""" from __future__ import annotations import logging from collections.abc import Sequence from pathlib import Path from typing import Any, Optional, Union from pydantic_ai.settings import ModelSettings from pipeline.protocol import SampleResult from ..core.context import ACEStepContext, SkillbookView from ..core.environments import Sample, TaskEnvironment from ..core.outputs import AgentOutput from ..core.skillbook import Skillbook from ..implementations import Agent, Reflector, SkillManager from ..integrations import wrap_skillbook_context from ..protocols import ( AgentLike, DeduplicationConfig, DeduplicationManagerLike, ReflectorLike, SkillManagerLike, ) from ..providers.pydantic_ai import resolve_model, settings_from_config from ..steps import learning_tail from .ace import ACE from .trace_analyser import TraceAnalyser logger = logging.getLogger(__name__) class ACELiteLLM: """PydanticAI-powered conversational agent with ACE learning. Bundles Agent, Reflector, SkillManager, and Skillbook into a simple interface. Delegates to :class:`ACE` for batch learning and :class:`TraceAnalyser` for trace-based learning. Two construction paths: 1. ``ACELiteLLM("gpt-4o-mini", ...)`` — builds PydanticAI-backed roles from a model string. 2. ``ACELiteLLM.from_model("gpt-4o-mini", ...)`` — explicit factory with full parameter control. Example:: ace = ACELiteLLM.from_model("gpt-4o-mini") answer = ace.ask("What is 2+2?") ace.learn(samples, environment=SimpleEnvironment(), epochs=3) ace.save("learned.json") """ def __init__( self, model: str, *, skillbook: Skillbook | None = None, skillbook_path: str | None = None, environment: TaskEnvironment | None = None, agent: AgentLike | None = None, reflector: ReflectorLike | None = None, skill_manager: SkillManagerLike | None = None, model_settings: ModelSettings | None = None, dedup_config: DeduplicationConfig | None = None, dedup_manager: DeduplicationManagerLike | None = None, dedup_interval: int = 10, checkpoint_dir: str | Path | None = None, checkpoint_interval: int = 10, is_learning: bool = True, logfire: bool = False, ) -> None: # Resolve skillbook if skillbook_path: self._skillbook = Skillbook.load_from_file(skillbook_path) elif skillbook is not None: self._skillbook = skillbook else: self._skillbook = Skillbook() # Build roles (use provided or create PydanticAI-backed defaults) self.agent: AgentLike = agent or Agent(model, model_settings=model_settings) self.reflector: ReflectorLike = reflector or Reflector( model, model_settings=model_settings ) self.skill_manager: SkillManagerLike = skill_manager or SkillManager( model, model_settings=model_settings ) self.environment = environment self.is_learning = is_learning # Resolve dedup manager if dedup_manager is not None: self._dedup_manager: DeduplicationManagerLike | None = dedup_manager elif dedup_config is not None: from ..deduplication import DeduplicationManager self._dedup_manager = DeduplicationManager(dedup_config) else: self._dedup_manager = None self._dedup_interval = dedup_interval self._checkpoint_dir = checkpoint_dir self._checkpoint_interval = checkpoint_interval # Logfire observability (explicit opt-in — fail loudly) if logfire: from ..observability import configure_logfire if not configure_logfire(): raise ImportError( "logfire=True requires the 'logfire' package. " "Install it with: pip install ace-framework[logfire]" ) # Lazy-init caches self._ace: ACE | None = None self._analyser: TraceAnalyser | None = None # Last interaction for learn_from_feedback() self._last_interaction: tuple[str, AgentOutput] | None = None # ------------------------------------------------------------------ # Alternative constructors # ------------------------------------------------------------------ @classmethod def from_setup( cls, *, config_dir: str | Path | None = None, validate: bool = False, **kwargs: Any, ) -> ACELiteLLM: """Build from ace.toml + .env (created by ``ace setup``). Looks for ``ace.toml`` in the current directory and parents. Loads ``.env`` for API keys. Optionally validates each model connection before proceeding. Args: config_dir: Explicit directory containing ace.toml. If None, searches current directory and parents. validate: If True, run a test LLM call for each configured model before building. Raises on failure. **kwargs: Extra kwargs forwarded to the constructor (skillbook, environment, etc.). Raises: FileNotFoundError: If no ace.toml is found. ConnectionError: If validate=True and a model fails. """ from ..providers.config import ( find_config, load_config, load_dotenv, ) # Load .env first so keys are available load_dotenv() # Find and load ace.toml if config_dir is not None: config = load_config(config_dir) else: config_path = find_config() if config_path is None: raise FileNotFoundError( "No ace.toml found. Run `ace setup` to create one, " "or use ACELiteLLM.from_model() / ACELiteLLM.from_config()." ) config = load_config(config_path.parent) return cls.from_config(config, validate=validate, **kwargs) @classmethod def from_config( cls, config: Any, # ACEModelConfig — Any to avoid circular import at module level *, validate: bool = False, **kwargs: Any, ) -> ACELiteLLM: """Build from an ``ACEModelConfig`` with per-role model selection. API keys are resolved from the environment (not from config). Each role gets its own PydanticAI-backed implementation, allowing different models for Agent, Reflector, and SkillManager. Args: config: An ``ACEModelConfig`` mapping roles to models. validate: If True, validate each model connection first. **kwargs: Extra kwargs forwarded to the constructor (skillbook, environment, etc.). Example:: from ace.providers.config import ACEModelConfig, ModelConfig config = ACEModelConfig( default=ModelConfig(model="gpt-4o-mini"), agent=ModelConfig(model="claude-sonnet-4-20250514"), ) ace = ACELiteLLM.from_config(config) """ from ..providers.config import ACEModelConfig if not isinstance(config, ACEModelConfig): raise TypeError(f"Expected ACEModelConfig, got {type(config).__name__}") if validate: from ..providers.registry import validate_connection seen: set[str] = set() for role in ("agent", "reflector", "skill_manager"): mc = config.for_role(role) if mc.model in seen: continue seen.add(mc.model) result = validate_connection(mc.model) if not result.success: raise ConnectionError( f"Model '{mc.model}' (for {role}) failed validation: " f"{result.error}" ) agent_config = config.for_role("agent") reflector_config = config.for_role("reflector") sm_config = config.for_role("skill_manager") return cls( agent_config.model, agent=Agent( agent_config.model, model_settings=settings_from_config(agent_config), ), reflector=Reflector( reflector_config.model, model_settings=settings_from_config(reflector_config), ), skill_manager=SkillManager( sm_config.model, model_settings=settings_from_config(sm_config), ), **kwargs, ) @classmethod def from_model( cls, model: str = "gpt-4o-mini", *, max_tokens: int = 2048, temperature: float = 0.0, skillbook: Skillbook | None = None, skillbook_path: Optional[str] = None, environment: Optional[TaskEnvironment] = None, dedup_config: Optional[DeduplicationConfig] = None, dedup_interval: int = 10, checkpoint_dir: Optional[Union[str, Path]] = None, checkpoint_interval: int = 10, is_learning: bool = True, logfire: bool = False, ) -> ACELiteLLM: """Build from a model string. Args: model: LiteLLM model identifier (e.g. ``"gpt-4o-mini"``). max_tokens: Max tokens for LLM responses. temperature: Sampling temperature. skillbook: Starting skillbook. skillbook_path: Path to load skillbook from. environment: Task environment for evaluation. dedup_config: Deduplication configuration. dedup_interval: Samples between deduplication runs. checkpoint_dir: Directory for checkpoint files. checkpoint_interval: Samples between checkpoint saves. is_learning: Whether learning is enabled. logfire: Enable Logfire observability (auto-instruments PydanticAI). """ model_settings = ModelSettings( temperature=temperature, max_tokens=max_tokens, ) return cls( model, model_settings=model_settings, skillbook=skillbook, skillbook_path=skillbook_path, environment=environment, dedup_config=dedup_config, dedup_interval=dedup_interval, checkpoint_dir=checkpoint_dir, checkpoint_interval=checkpoint_interval, is_learning=is_learning, logfire=logfire, ) # ------------------------------------------------------------------ # Lazy-init runners # ------------------------------------------------------------------ def _get_extra_steps(self) -> list[Any] | None: """Return extra pipeline steps or None.""" return None def _get_ace(self, environment: TaskEnvironment | None = None) -> ACE: """Return (or build) the cached ACE runner.""" env = environment or self.environment # Invalidate if environment changed if self._ace is not None and env is not self.environment: self._ace = None self.environment = env if self._ace is None: self._ace = ACE.from_roles( agent=self.agent, reflector=self.reflector, skill_manager=self.skill_manager, environment=env, skillbook=self._skillbook, dedup_manager=self._dedup_manager, dedup_interval=self._dedup_interval, checkpoint_dir=self._checkpoint_dir, checkpoint_interval=self._checkpoint_interval, extra_steps=self._get_extra_steps(), ) return self._ace def _get_analyser(self) -> TraceAnalyser: """Return (or build) the cached TraceAnalyser.""" if self._analyser is None: self._analyser = TraceAnalyser.from_roles( reflector=self.reflector, skill_manager=self.skill_manager, skillbook=self._skillbook, dedup_manager=self._dedup_manager, dedup_interval=self._dedup_interval, checkpoint_dir=self._checkpoint_dir, checkpoint_interval=self._checkpoint_interval, extra_steps=self._get_extra_steps(), ) return self._analyser # ------------------------------------------------------------------ # Public API # ------------------------------------------------------------------ def ask(self, question: str, context: str = "") -> str: """Ask a question using the current skillbook. Direct Agent call — does not go through the pipeline. Stores the interaction for optional :meth:`learn_from_feedback`. Args: question: The question to answer. context: Optional context for the question. Returns: The agent's final answer. """ output = self.agent.generate( question=question, context=context, skillbook=self._skillbook, ) self._last_interaction = (question, output) return output.final_answer def learn( self, samples: Sequence[Sample], environment: TaskEnvironment | None = None, epochs: int = 1, *, wait: bool = True, ) -> list[SampleResult]: """Run the full ACE learning pipeline over samples. Args: samples: Training samples with questions and ground truth. environment: Task environment for evaluation. Falls back to the environment set at construction time. epochs: Number of passes over the samples. wait: If ``True``, block until background learning completes. Returns: List of ``SampleResult``, one per sample per epoch. Raises: RuntimeError: If learning is disabled. """ if not self.is_learning: raise RuntimeError("Learning is disabled. Call enable_learning() first.") return self._get_ace(environment).run(samples, epochs=epochs, wait=wait) def learn_from_traces( self, traces: Sequence[Any], epochs: int = 1, *, wait: bool = True, ) -> list[SampleResult]: """Learn from pre-recorded execution traces. Args: traces: Raw trace objects (dicts, framework results, etc.). epochs: Number of passes over the traces. wait: If ``True``, block until background learning completes. Returns: List of ``SampleResult``, one per trace per epoch. Raises: RuntimeError: If learning is disabled. """ if not self.is_learning: raise RuntimeError("Learning is disabled. Call enable_learning() first.") return self._get_analyser().run(traces, epochs=epochs, wait=wait) def learn_from_feedback( self, feedback: str, ground_truth: str | None = None, ) -> bool: """Learn from the last :meth:`ask` interaction. Runs the standard ``learning_tail`` pipeline (ReflectStep, UpdateStep) on the most recent ``ask()`` call with the provided feedback. The agentic SkillManager mutates the skillbook directly — there is no separate apply step. Args: feedback: User feedback about the answer quality. ground_truth: Optional correct answer. Returns: ``True`` if learning was applied, ``False`` if no prior interaction exists or learning is disabled. """ if not self.is_learning or self._last_interaction is None: return False question, agent_output = self._last_interaction # Build synthetic trace (same format EvaluateStep produces) trace = { "question": question, "context": "", "ground_truth": ground_truth, "reasoning": agent_output.reasoning, "answer": agent_output.final_answer, "skill_ids": agent_output.skill_ids, "feedback": feedback, } ctx = ACEStepContext( skillbook=SkillbookView(self._skillbook), trace=trace, ) # Same learning pipeline as learn() and learn_from_traces() from pipeline import Pipeline steps = learning_tail( self.reflector, self.skill_manager, self._skillbook, ) pipe = Pipeline(steps) results = pipe.run([ctx]) pipe.wait_for_background() # ReflectStep has async_boundary=True return len(results) > 0 and results[0].error is None # ------------------------------------------------------------------ # Lifecycle # ------------------------------------------------------------------ @property def skillbook(self) -> Skillbook: """The current skillbook.""" return self._skillbook def save(self, path: str) -> None: """Save the skillbook to disk.""" self._skillbook.save_to_file(path) def load(self, path: str) -> None: """Load a skillbook from disk. Invalidates cached runners (they hold stale skillbook refs). """ self._skillbook = Skillbook.load_from_file(path) self._ace = None self._analyser = None def enable_learning(self) -> None: """Enable learning.""" self.is_learning = True def disable_learning(self) -> None: """Disable learning.""" self.is_learning = False def get_strategies(self) -> str: """Return formatted skillbook strategies for display.""" if not self._skillbook.skills(): return "" return wrap_skillbook_context(self._skillbook) def wait_for_background(self, timeout: float | None = None) -> None: """Block until all background learning completes.""" if self._ace is not None: self._ace.wait_for_background(timeout) if self._analyser is not None: self._analyser.wait_for_background(timeout) @property def learning_stats(self) -> dict[str, int]: """Return background learning progress.""" stats: dict[str, int] = {} if self._ace is not None: stats.update(self._ace.learning_stats) if self._analyser is not None: stats.update(self._analyser.learning_stats) return stats # Backward-compat aliases save_skillbook = save load_skillbook = load wait_for_learning = wait_for_background