lm-deluge 0.0.67__py3-none-any.whl → 0.0.90__py3-none-any.whl

lm_deluge/pipelines/gepa/optimizer.py

@@ -0,0 +1,435 @@
+"""
+GEPA optimizer.
+
+Main optimization loop that evolves text components using trajectory-based feedback.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import random
+from collections.abc import Awaitable, Callable
+from pathlib import Path
+from typing import Generic, TypeVar
+
+from lm_deluge.client import _LLMClient
+from lm_deluge.pipelines.gepa.core import (
+    Component,
+    EvalResult,
+    GEPAResult,
+    GEPAState,
+)
+from lm_deluge.pipelines.gepa.proposer import propose_improvement
+
+T = TypeVar("T")  # Example type
+
+# Type for the user-provided evaluate function (must be async)
+AsyncEvaluateFn = Callable[[_LLMClient, dict[str, str], T], Awaitable[EvalResult]]
+
+
+def optimize(
+    components: dict[str, Component],
+    evaluate_fn: AsyncEvaluateFn[T],
+    dataset: list[T],
+    task_client: _LLMClient,
+    proposer_client: _LLMClient,
+    *,
+    val_dataset: list[T] | None = None,
+    max_iterations: int = 100,
+    max_evals: int | None = None,
+    minibatch_size: int = 4,
+    perfect_score: float = 1.0,
+    proposal_prompt_template: str | None = None,
+    meta_instructions: str | None = None,
+    run_dir: str | Path | None = None,
+    log_fn: Callable[[str], None] | None = None,
+    save_trajectories: bool = False,
+    seed: int = 0,
+) -> GEPAResult:
+    """
+    Run GEPA optimization to improve text components.
+
+    Args:
+        components: The text components to optimize (name -> Component)
+        evaluate_fn: Async function that evaluates one example: (client, values, example) -> EvalResult
+        dataset: Training examples
+        task_client: LLMClient for running evaluations
+        proposer_client: LLMClient for generating proposals
+        val_dataset: Optional separate validation set (defaults to dataset)
+        max_iterations: Maximum optimization iterations
+        max_evals: Maximum total evaluations (budget)
+        minibatch_size: Examples per minibatch for proposal evaluation
+        perfect_score: Score considered perfect (skip examples that achieve this)
+        proposal_prompt_template: Custom prompt template for proposer
+        meta_instructions: Guidelines to steer the proposer (e.g., "Don't overfit to specific examples")
+        run_dir: Directory to save state and trajectories
+        log_fn: Logging function (defaults to print)
+        save_trajectories: Whether to save trajectories to disk
+        seed: Random seed for reproducibility
+
+    Returns:
+        GEPAResult with optimization results
+    """
+    engine = GEPAEngine(
+        components=components,
+        evaluate_fn=evaluate_fn,
+        dataset=dataset,
+        task_client=task_client,
+        proposer_client=proposer_client,
+        val_dataset=val_dataset,
+        max_iterations=max_iterations,
+        max_evals=max_evals,
+        minibatch_size=minibatch_size,
+        perfect_score=perfect_score,
+        proposal_prompt_template=proposal_prompt_template,
+        meta_instructions=meta_instructions,
+        run_dir=run_dir,
+        log_fn=log_fn,
+        save_trajectories=save_trajectories,
+        seed=seed,
+    )
+
+    return asyncio.run(engine.run())
+
+
+class GEPAEngine(Generic[T]):
+    """
+    Stateful GEPA optimizer.
+
+    Use this for more control over the optimization process:
+    - Resume from saved state
+    - Step through iterations manually
+    - Access intermediate state
+    """
+
+    def __init__(
+        self,
+        components: dict[str, Component],
+        evaluate_fn: AsyncEvaluateFn[T],
+        dataset: list[T],
+        task_client: _LLMClient,
+        proposer_client: _LLMClient,
+        *,
+        val_dataset: list[T] | None = None,
+        max_iterations: int = 100,
+        max_evals: int | None = None,
+        minibatch_size: int = 4,
+        perfect_score: float = 1.0,
+        proposal_prompt_template: str | None = None,
+        meta_instructions: str | None = None,
+        run_dir: str | Path | None = None,
+        log_fn: Callable[[str], None] | None = None,
+        save_trajectories: bool = False,
+        seed: int = 0,
+    ):
+        self.components = components
+        self.evaluate_fn = evaluate_fn
+        self.dataset = dataset
+        self.task_client = task_client
+        self.proposer_client = proposer_client
+        self.val_dataset = val_dataset if val_dataset is not None else dataset
+        self.max_iterations = max_iterations
+        self.max_evals = max_evals
+        self.minibatch_size = minibatch_size
+        self.perfect_score = perfect_score
+        self.proposal_prompt_template = proposal_prompt_template
+        self.meta_instructions = meta_instructions
+        self.run_dir = Path(run_dir) if run_dir else None
+        self.log_fn = log_fn or print
+        self.save_trajectories = save_trajectories
+
+        self.rng = random.Random(seed)
+        self.state: GEPAState | None = None
+        self._trajectory_counter = 0
+
+    def _log(self, msg: str) -> None:
+        self.log_fn(msg)
+
+    async def _evaluate_batch(
+        self,
+        examples: list[tuple[int, T]],  # (index, example) pairs
+        component_values: dict[str, str],
+    ) -> list[tuple[int, EvalResult]]:
+        """Evaluate a batch of examples concurrently, return (index, result) pairs."""
+
+        async def eval_one(idx: int, example: T) -> tuple[int, EvalResult] | None:
+            try:
+                result = await self.evaluate_fn(
+                    self.task_client, component_values, example
+                )
+                return (idx, result)
+            except Exception as e:
+                self._log(f"Error evaluating example {idx}: {e}")
+                return None
+
+        # Run all evaluations concurrently
+        tasks = [eval_one(idx, example) for idx, example in examples]
+        results_raw = await asyncio.gather(*tasks)
+
+        # Filter out None results (failed evaluations)
+        results = [r for r in results_raw if r is not None]
+        return results
+
+    async def _evaluate_all(
+        self,
+        examples: list[T],
+        component_values: dict[str, str],
+    ) -> dict[int, float]:
+        """Evaluate all examples, return scores dict."""
+        indexed = [(i, ex) for i, ex in enumerate(examples)]
+        results = await self._evaluate_batch(indexed, component_values)
+
+        if self.state:
+            self.state.total_evals += len(results)
+
+        return {idx: result.score for idx, result in results}
+
+    def _save_trajectory(
+        self,
+        iteration: int,
+        tag: str,
+        candidate_values: dict[str, str],
+        example_idx: int,
+        result: EvalResult,
+    ) -> None:
+        """Save a trajectory to disk for debugging."""
+        if not self.save_trajectories or not self.run_dir:
+            return
+
+        traj_dir = self.run_dir / "trajectories"
+        traj_dir.mkdir(parents=True, exist_ok=True)
+
+        self._trajectory_counter += 1
+        filename = f"{self._trajectory_counter:04d}_iter{iteration}_{tag}.json"
+
+        data = {
+            "iteration": iteration,
+            "tag": tag,
+            "example_idx": example_idx,
+            "candidate": candidate_values,
+            "score": result.score,
+            "feedback": result.feedback,
+            "conversation": result.conversation.to_log(),
+        }
+
+        try:
+            with open(traj_dir / filename, "w") as f:
+                json.dump(data, f, indent=2, default=str)
+        except Exception as e:
+            self._log(f"Failed to save trajectory: {e}")
+
+    async def initialize(self) -> None:
+        """Initialize state by evaluating seed candidate on validation set."""
+        self._log("Evaluating seed candidate...")
+
+        seed_values = {name: comp.value for name, comp in self.components.items()}
+        seed_scores = await self._evaluate_all(self.val_dataset, seed_values)
+
+        self.state = GEPAState.initialize(self.components, seed_scores)
+
+        avg_score = sum(seed_scores.values()) / len(seed_scores) if seed_scores else 0.0
+        self._log(
+            f"Seed candidate: avg_score={avg_score:.4f} on {len(seed_scores)} examples"
+        )
+
+    def _should_stop(self) -> bool:
+        """Check if we should stop optimization."""
+        if self.state is None:
+            return True
+
+        if self.state.iteration >= self.max_iterations:
+            return True
+
+        if self.max_evals and self.state.total_evals >= self.max_evals:
+            return True
+
+        # Stop if best candidate achieves perfect score on val set
+        best_idx = self.state.best_candidate_idx()
+        best_avg = self.state.get_candidate_avg_score(best_idx)
+        if best_avg >= self.perfect_score:
+            self._log(
+                f"Best candidate achieved perfect score ({best_avg:.4f}), stopping"
+            )
+            return True
+
+        return False
+
+    async def step(self) -> bool:
+        """
+        Run one optimization iteration.
+
+        Returns:
+            True if optimization should continue, False if done
+        """
+        if self.state is None:
+            await self.initialize()
+
+        assert self.state is not None
+
+        if self._should_stop():
+            return False
+
+        self.state.iteration += 1
+        iteration = self.state.iteration
+
+        # Get current best candidate
+        best_idx = self.state.best_candidate_idx()
+        current_values = self.state.candidates[best_idx]
+
+        # Find examples where the BEST candidate isn't perfect
+        # (not Pareto front - we want to improve the best single candidate)
+        best_scores = self.state.candidate_scores[best_idx]
+        improvable = [
+            ex_idx
+            for ex_idx, score in best_scores.items()
+            if score < self.perfect_score
+        ]
+
+        if not improvable:
+            # Best candidate is perfect on all examples it was evaluated on
+            # Just pick a random example to re-evaluate and potentially find issues
+            improvable = list(best_scores.keys())
+            if not improvable:
+                self._log(f"Iteration {iteration}: No examples to evaluate")
+                return False
+
+        # Pick an example to focus on (prefer non-perfect ones)
+        focus_idx = self.rng.choice(improvable)
+
+        # Evaluate current candidate on focus example to get trajectory
+        focus_example = self.val_dataset[focus_idx]
+        results = await self._evaluate_batch(
+            [(focus_idx, focus_example)], current_values
+        )
+
+        if not results:
+            self._log(f"Iteration {iteration}: Failed to evaluate focus example")
+            return True
+
+        _, focus_result = results[0]
+        self.state.total_evals += 1
+
+        if self.save_trajectories:
+            self._save_trajectory(
+                iteration, "focus", current_values, focus_idx, focus_result
+            )
+
+        # Generate proposal
+        proposal = await propose_improvement(
+            proposer_client=self.proposer_client,
+            eval_result=focus_result,
+            components=self.components,
+            current_values=current_values,
+            prompt_template=self.proposal_prompt_template,
+            meta_instructions=self.meta_instructions,
+        )
+
+        if proposal is None:
+            self._log(f"Iteration {iteration}: No proposal generated")
+            return True
+
+        self._log(
+            f"Iteration {iteration}: Proposing change to '{proposal.component_name}' - {proposal.reasoning[:80]}..."
+        )
+
+        # Build new candidate
+        new_values = dict(current_values)
+        new_values[proposal.component_name] = proposal.new_value
+
+        # Evaluate on minibatch (including focus example)
+        minibatch_indices = [focus_idx]
+        other_indices = [i for i in improvable if i != focus_idx]
+        if other_indices:
+            additional = self.rng.sample(
+                other_indices, min(self.minibatch_size - 1, len(other_indices))
+            )
+            minibatch_indices.extend(additional)
+
+        # Evaluate old and new candidates on minibatch concurrently
+        minibatch = [(i, self.val_dataset[i]) for i in minibatch_indices]
+        old_results, new_results = await asyncio.gather(
+            self._evaluate_batch(minibatch, current_values),
+            self._evaluate_batch(minibatch, new_values),
+        )
+
+        old_sum = sum(r.score for _, r in old_results)
+        new_sum = sum(r.score for _, r in new_results)
+
+        self.state.total_evals += len(old_results) + len(new_results)
+
+        # Accept if improved
+        if new_sum <= old_sum:
+            self._log(
+                f"Iteration {iteration}: Rejected (old={old_sum:.3f}, new={new_sum:.3f})"
+            )
+            return True
+
+        self._log(
+            f"Iteration {iteration}: Accepted (old={old_sum:.3f}, new={new_sum:.3f})"
+        )
+
+        # Full validation evaluation
+        val_scores = await self._evaluate_all(self.val_dataset, new_values)
+
+        # Add to population
+        new_idx = self.state.add_candidate(new_values, best_idx, val_scores)
+
+        new_avg = sum(val_scores.values()) / len(val_scores) if val_scores else 0.0
+        best_avg = self.state.get_candidate_avg_score(self.state.best_candidate_idx())
+        self._log(
+            f"  New candidate {new_idx}: val_avg={new_avg:.4f}, best={best_avg:.4f}, "
+            f"pool={len(self.state.candidates)}"
+        )
+
+        # Save state periodically
+        if self.run_dir and iteration % 10 == 0:
+            self.state.save(self.run_dir)
+
+        return True
+
+    async def run(self) -> GEPAResult:
+        """Run optimization until stopping condition."""
+        if self.state is None:
+            await self.initialize()
+
+        try:
+            from tqdm import tqdm
+
+            pbar = tqdm(
+                total=self.max_iterations,
+                desc="GEPA",
+                unit="iter",
+            )
+        except ImportError:
+            pbar = None
+
+        while await self.step():
+            if pbar:
+                pbar.update(1)
+                pbar.set_postfix(
+                    evals=self.state.total_evals if self.state else 0,
+                    best=f"{self.state.get_candidate_avg_score(self.state.best_candidate_idx()):.3f}"
+                    if self.state
+                    else 0,
+                )
+
+        if pbar:
+            pbar.close()
+
+        # Save final state
+        if self.run_dir and self.state:
+            self.state.save(self.run_dir)
+
+        return self.result()
+
+    def result(self) -> GEPAResult:
+        """Get current result as immutable snapshot."""
+        if self.state is None:
+            raise RuntimeError(
+                "Optimizer not initialized. Call initialize() or run() first."
+            )
+
+        return GEPAResult.from_state(
+            self.state, run_dir=str(self.run_dir) if self.run_dir else None
+        )

lm_deluge/pipelines/gepa/proposer.py

@@ -0,0 +1,235 @@
+"""
+Proposer for GEPA optimization.
+
+The proposer analyzes a trajectory and proposes improvements to ONE component.
+"""
+
+from __future__ import annotations
+
+from lm_deluge.prompt import Conversation
+
+from lm_deluge.pipelines.gepa.core import Component, EvalResult, Proposal
+from lm_deluge.pipelines.gepa.util import (
+    extract_text_from_response,
+    format_components_for_prompt,
+    format_conversation_compact,
+)
+
+
+DEFAULT_PROPOSAL_PROMPT = """You are optimizing an AI system by improving its text configuration.
+
+## The Trajectory
+
+Below is a conversation showing what the AI did on a task:
+
+<trajectory>
+{trajectory}
+</trajectory>
+
+## Feedback
+
+{feedback}
+
+## Components
+
+These are the text components that control the AI's behavior. You can modify ONE of them:
+
+{components}
+{meta_instructions}
+## Your Task
+
+1. Analyze the trajectory to understand what went wrong (or could be better)
+2. Identify which component is most likely responsible
+3. Propose a specific improvement to that ONE component
+
+Think about:
+- Did the AI misunderstand the task? (maybe the system prompt needs clarity)
+- Did it use tools incorrectly? (maybe tool descriptions need improvement)
+- Did it miss important information? (maybe instructions need to be more explicit)
+
+## Response Format
+
+Respond with:
+COMPONENT: <name of the component to change>
+REASONING: <1-2 sentences on why this change will help>
+NEW_VALUE:
+```
+<the complete improved text for this component>
+```
+"""
+
+
+def build_proposal_prompt(
+    conversation: Conversation,
+    feedback: str,
+    components: dict[str, Component],
+    current_values: dict[str, str],
+    template: str | None = None,
+    meta_instructions: str | None = None,
+) -> str:
+    """
+    Build the prompt for the proposer LLM.
+
+    Args:
+        conversation: The trajectory to analyze
+        feedback: Feedback on the result
+        components: Component definitions (with descriptions)
+        current_values: Current text values for each component
+        template: Optional custom prompt template
+        meta_instructions: Optional instructions to guide the proposer's behavior
+                          (e.g., "Focus on general improvements, don't overfit to specific examples")
+
+    Returns:
+        Formatted prompt string
+    """
+    template = template or DEFAULT_PROPOSAL_PROMPT
+
+    # Format trajectory
+    trajectory_str = format_conversation_compact(conversation)
+
+    # Format components
+    descriptions = {name: comp.description for name, comp in components.items()}
+    components_str = format_components_for_prompt(current_values, descriptions)
+
+    # Format meta instructions
+    if meta_instructions:
+        meta_str = f"\n## Guidelines\n\n{meta_instructions}\n\n"
+    else:
+        meta_str = "\n"
+
+    return template.format(
+        trajectory=trajectory_str,
+        feedback=feedback,
+        components=components_str,
+        meta_instructions=meta_str,
+    )
+
+
+def parse_proposal_response(
+    response: str, valid_components: list[str]
+) -> Proposal | None:
+    """
+    Parse the proposer's response to extract the proposal.
+
+    Args:
+        response: Raw LLM response
+        valid_components: List of valid component names
+
+    Returns:
+        Proposal if parsing succeeded, None otherwise
+    """
+    # Find COMPONENT line
+    component_name = None
+    for line in response.split("\n"):
+        line = line.strip()
+        if line.upper().startswith("COMPONENT:"):
+            component_name = line.split(":", 1)[1].strip()
+            break
+
+    if not component_name:
+        return None
+
+    # Validate component name
+    if component_name not in valid_components:
+        # Try case-insensitive match
+        for valid in valid_components:
+            if valid.lower() == component_name.lower():
+                component_name = valid
+                break
+        else:
+            return None
+
+    # Find REASONING line
+    reasoning = ""
+    for line in response.split("\n"):
+        line = line.strip()
+        if line.upper().startswith("REASONING:"):
+            reasoning = line.split(":", 1)[1].strip()
+            break
+
+    # Extract new value from code block
+    new_value = extract_text_from_response(response)
+    if not new_value:
+        return None
+
+    return Proposal(
+        component_name=component_name,
+        new_value=new_value,
+        reasoning=reasoning,
+    )
+
+
+async def propose_improvement(
+    proposer_client,  # LLMClient
+    eval_result: EvalResult,
+    components: dict[str, Component],
+    current_values: dict[str, str],
+    prompt_template: str | None = None,
+    meta_instructions: str | None = None,
+) -> Proposal | None:
+    """
+    Use an LLM to propose an improvement to one component.
+
+    Args:
+        proposer_client: LLMClient for generating proposals
+        eval_result: The evaluation result containing trajectory and feedback
+        components: Component definitions
+        current_values: Current text values
+        prompt_template: Optional custom prompt template
+        meta_instructions: Optional guidelines to steer the proposer
+                          (e.g., "Don't overfit to specific examples")
+
+    Returns:
+        Proposal if successful, None otherwise
+    """
+    # Build prompt
+    prompt = build_proposal_prompt(
+        conversation=eval_result.conversation,
+        feedback=eval_result.feedback,
+        components=components,
+        current_values=current_values,
+        template=prompt_template,
+        meta_instructions=meta_instructions,
+    )
+
+    # Call LLM
+    response = await proposer_client.start(prompt)
+    if not response or not response.completion:
+        return None
+
+    response_text = response.completion
+
+    # Parse response
+    valid_components = list(components.keys())
+    return parse_proposal_response(response_text, valid_components)
+
+
+def propose_improvement_sync(
+    proposer_client,  # LLMClient
+    eval_result: EvalResult,
+    components: dict[str, Component],
+    current_values: dict[str, str],
+    prompt_template: str | None = None,
+) -> Proposal | None:
+    """
+    Synchronous version of propose_improvement.
+    """
+    # Build prompt
+    prompt = build_proposal_prompt(
+        conversation=eval_result.conversation,
+        feedback=eval_result.feedback,
+        components=components,
+        current_values=current_values,
+        template=prompt_template,
+    )
+
+    # Call LLM
+    responses = proposer_client.process_prompts_sync([prompt], show_progress=False)
+    if not responses or not responses[0].completion:
+        return None
+
+    response_text = responses[0].completion
+
+    # Parse response
+    valid_components = list(components.keys())
+    return parse_proposal_response(response_text, valid_components)