opik-optimizer 1.0.6__py3-none-any.whl → 1.1.0__py3-none-any.whl

opik_optimizer/evolutionary_optimizer/helpers.py

@@ -0,0 +1,10 @@
+from ..optimization_config import chat_prompt
+
+
+class Helpers:
+    def _get_task_description_for_llm(self, prompt: chat_prompt.ChatPrompt) -> str:
+        """Generates a concise task description for LLM prompts that need context."""
+        description = "Task: Given a list of AI messages with placeholder values, generate an effective prompt. "
+        description += f"The original high-level instruction being optimized is: '{prompt.get_messages()}'. "
+        description += "The goal is to create an effective prompt that guides a language model to perform this task well."
+        return description

opik_optimizer/evolutionary_optimizer/llm_support.py

@@ -0,0 +1,134 @@
+from typing import Any, TYPE_CHECKING
+
+import logging
+import os
+import time
+import random
+
+import litellm
+from litellm import exceptions as litellm_exceptions
+from litellm.caching import Cache
+from litellm.types.caching import LiteLLMCacheType
+from opik.evaluation.models.litellm import opik_monitor as opik_litellm_monitor
+
+from .. import _throttle
+
+
+logger = logging.getLogger(__name__)
+
+
+# Configure LiteLLM cache with safe fallback
+try:
+    # Prefer a disk cache in a user-writable location
+    cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "litellm")
+    os.makedirs(cache_dir, exist_ok=True)
+    litellm.cache = Cache(type=LiteLLMCacheType.DISK, cache_dir=cache_dir)
+except (PermissionError, OSError, FileNotFoundError):
+    # Fall back to in-memory cache to avoid disk timeouts/locks
+    litellm.cache = Cache(type=LiteLLMCacheType.MEMORY)
+
+_rate_limiter = _throttle.get_rate_limiter_for_current_opik_installation()
+
+
+class LlmSupport:
+    if TYPE_CHECKING:
+        model: str
+        llm_call_counter: int
+        project_name: str | None
+        disable_litellm_monitoring: bool
+        temperature: float
+        max_tokens: int
+        top_p: float
+        frequency_penalty: float
+        presence_penalty: float
+
+    @_throttle.rate_limited(_rate_limiter)
+    def _call_model(
+        self,
+        messages: list[dict[str, str]],
+        is_reasoning: bool = False,
+        optimization_id: str | None = None,
+    ) -> str:
+        """Call the model with the given prompt and return the response string."""
+        # Build base call params
+        llm_config_params: dict[str, Any] = {
+            "temperature": getattr(self, "temperature", 0.3),
+            "max_tokens": getattr(self, "max_tokens", 1000),
+            "top_p": getattr(self, "top_p", 1.0),
+            "frequency_penalty": getattr(self, "frequency_penalty", 0.0),
+            "presence_penalty": getattr(self, "presence_penalty", 0.0),
+        }
+
+        # Add Opik metadata unless disabled
+        try:
+            disable_monitoring_env = os.getenv(
+                "OPIK_OPTIMIZER_DISABLE_LITELLM_MONITORING", "0"
+            )
+            disable_monitoring = getattr(
+                self, "disable_litellm_monitoring", False
+            ) or disable_monitoring_env.lower() in ("1", "true", "yes")
+
+            if not disable_monitoring:
+                metadata_for_opik: dict[str, Any] = {}
+                pn = getattr(self, "project_name", None)
+                if pn:
+                    metadata_for_opik["project_name"] = pn
+                    metadata_for_opik["opik"] = {"project_name": pn}
+                if optimization_id and "opik" in metadata_for_opik:
+                    metadata_for_opik["opik"]["optimization_id"] = optimization_id
+                metadata_for_opik["optimizer_name"] = self.__class__.__name__
+                metadata_for_opik["opik_call_type"] = (
+                    "reasoning" if is_reasoning else "evaluation_llm_task_direct"
+                )
+                if metadata_for_opik:
+                    llm_config_params["metadata"] = metadata_for_opik
+
+                # Try to add Opik monitoring callbacks; fall back silently on failure
+                llm_config_params = (
+                    opik_litellm_monitor.try_add_opik_monitoring_to_params(  # type: ignore
+                        llm_config_params.copy()
+                    )
+                )
+        except Exception as e:
+            logger.debug(f"Skipping Opik-LiteLLM monitoring setup: {e}")
+
+        # Retry policy for transient errors
+        max_retries = int(os.getenv("OPIK_OPTIMIZER_LITELLM_MAX_RETRIES", "3"))
+        base_sleep = float(os.getenv("OPIK_OPTIMIZER_LITELLM_BACKOFF", "0.5"))
+
+        for attempt in range(max_retries + 1):
+            try:
+                logger.debug(
+                    f"Calling model '{self.model}' with messages: {messages}, params: {llm_config_params} (attempt {attempt + 1})"
+                )
+                response = litellm.completion(
+                    model=self.model, messages=messages, **llm_config_params
+                )
+                self.llm_call_counter += 1
+                return response.choices[0].message.content
+            except (
+                litellm_exceptions.RateLimitError,
+                litellm_exceptions.APIConnectionError,
+                litellm_exceptions.InternalServerError,
+            ) as e:
+                if attempt < max_retries:
+                    sleep_s = min(10.0, base_sleep * (2**attempt)) + random.uniform(
+                        0, 0.25
+                    )
+                    logger.warning(
+                        f"LiteLLM transient error ({type(e).__name__}): {e}. Retrying in {sleep_s:.2f}s..."
+                    )
+                    time.sleep(sleep_s)
+                    continue
+                logger.error(f"LiteLLM error (final attempt): {e}")
+                raise
+            except litellm_exceptions.ContextWindowExceededError as e:
+                logger.error(f"LiteLLM Context Window Exceeded Error: {e}")
+                raise
+            except Exception as e:
+                logger.error(
+                    f"Error calling model '{self.model}': {type(e).__name__} - {e}"
+                )
+                raise
+        # Should never reach here
+        raise RuntimeError("LLM call did not return a response and did not raise")

opik_optimizer/evolutionary_optimizer/mutation_ops.py

@@ -0,0 +1,292 @@
+from typing import Any, TYPE_CHECKING
+
+import json
+import logging
+import random
+
+from . import prompts as evo_prompts
+from ..optimization_config import chat_prompt
+from .. import utils
+from . import reporting
+
+
+logger = logging.getLogger(__name__)
+
+
+class MutationOps:
+    if TYPE_CHECKING:
+        _calculate_population_diversity: Any
+        DEFAULT_DIVERSITY_THRESHOLD: float
+        verbose: int
+        output_style_guidance: str
+        _get_task_description_for_llm: Any
+        _call_model: Any
+
+    def _deap_mutation(
+        self, individual: Any, initial_prompt: chat_prompt.ChatPrompt
+    ) -> Any:
+        """Enhanced mutation operation with multiple strategies."""
+        prompt = chat_prompt.ChatPrompt(messages=individual)
+
+        # Choose mutation strategy based on current diversity
+        diversity = self._calculate_population_diversity()
+
+        # Determine thresholds based on diversity
+        if diversity < self.DEFAULT_DIVERSITY_THRESHOLD:
+            # Low diversity - use more aggressive mutations (higher chance for semantic)
+            semantic_threshold = 0.5
+            structural_threshold = 0.8  # semantic_threshold + 0.3
+        else:
+            # Good diversity - use more conservative mutations (higher chance for word_level)
+            semantic_threshold = 0.4
+            structural_threshold = 0.7  # semantic_threshold + 0.3
+
+        mutation_choice = random.random()
+
+        if mutation_choice > structural_threshold:
+            mutated_prompt = self._word_level_mutation_prompt(prompt)
+            reporting.display_success(
+                "      Mutation successful, prompt has been edited by randomizing words (word-level mutation).",
+                verbose=self.verbose,
+            )
+            return type(individual)(mutated_prompt.get_messages())
+        elif mutation_choice > semantic_threshold:
+            mutated_prompt = self._structural_mutation(prompt)
+            reporting.display_success(
+                "      Mutation successful, prompt has been edited by reordering, combining, or splitting sentences (structural mutation).",
+                verbose=self.verbose,
+            )
+            return type(individual)(mutated_prompt.get_messages())
+        else:
+            mutated_prompt = self._semantic_mutation(prompt, initial_prompt)
+            reporting.display_success(
+                "      Mutation successful, prompt has been edited using an LLM (semantic mutation).",
+                verbose=self.verbose,
+            )
+            return type(individual)(mutated_prompt.get_messages())
+
+    def _semantic_mutation(
+        self, prompt: chat_prompt.ChatPrompt, initial_prompt: chat_prompt.ChatPrompt
+    ) -> chat_prompt.ChatPrompt:
+        """Enhanced semantic mutation with multiple strategies."""
+        current_output_style_guidance = self.output_style_guidance
+        if random.random() < 0.1:
+            return self._radical_innovation_mutation(prompt, initial_prompt)
+
+        try:
+            strategy = random.choice(
+                [
+                    "rephrase",
+                    "simplify",
+                    "elaborate",
+                    "restructure",
+                    "focus",
+                    "increase_complexity_and_detail",
+                ]
+            )
+
+            strategy_prompts = evo_prompts.mutation_strategy_prompts(
+                current_output_style_guidance
+            )
+            user_prompt_for_semantic_mutation = (
+                evo_prompts.semantic_mutation_user_prompt(
+                    prompt.get_messages(),
+                    self._get_task_description_for_llm(initial_prompt),
+                    current_output_style_guidance,
+                    strategy_prompts[strategy],
+                )
+            )
+            response = self._call_model(
+                messages=[
+                    {
+                        "role": "system",
+                        "content": evo_prompts.semantic_mutation_system_prompt(
+                            current_output_style_guidance
+                        ),
+                    },
+                    {"role": "user", "content": user_prompt_for_semantic_mutation},
+                ],
+                is_reasoning=True,
+            )
+
+            try:
+                messages = utils.json_to_dict(response.strip())
+            except Exception as parse_exc:
+                raise RuntimeError(
+                    f"Error parsing semantic mutation response as JSON. "
+                    f"Response: {response!r}\nOriginal error: {parse_exc}"
+                ) from parse_exc
+            return chat_prompt.ChatPrompt(messages=messages)
+        except Exception as e:
+            reporting.display_error(
+                f"      Error in semantic mutation, this is usually a parsing error: {e}",
+                verbose=self.verbose,
+            )
+            return prompt
+
+    def _structural_mutation(
+        self, prompt: chat_prompt.ChatPrompt
+    ) -> chat_prompt.ChatPrompt:
+        """Perform structural mutation (reordering, combining, splitting)."""
+        mutated_messages: list[dict[str, str]] = []
+
+        for message in prompt.get_messages():
+            content = message["content"]
+            role = message["role"]
+
+            sentences = [s.strip() for s in content.split(".") if s.strip()]
+            if len(sentences) <= 1:
+                mutated_messages.append(
+                    {"role": role, "content": self._word_level_mutation(content)}
+                )
+                continue
+
+            mutation_type = random.random()
+            if mutation_type < 0.3:
+                random.shuffle(sentences)
+                mutated_messages.append(
+                    {"role": role, "content": ". ".join(sentences) + "."}
+                )
+                continue
+            elif mutation_type < 0.6:
+                if len(sentences) >= 2:
+                    idx = random.randint(0, len(sentences) - 2)
+                    combined = sentences[idx] + " and " + sentences[idx + 1]
+                    sentences[idx : idx + 2] = [combined]
+                    mutated_messages.append(
+                        {"role": role, "content": ". ".join(sentences) + "."}
+                    )
+                    continue
+            else:
+                idx = random.randint(0, len(sentences) - 1)
+                words = sentences[idx].split()
+                if len(words) > 3:
+                    split_point = random.randint(2, len(words) - 2)
+                    sentences[idx : idx + 1] = [
+                        " ".join(words[:split_point]),
+                        " ".join(words[split_point:]),
+                    ]
+                    mutated_messages.append(
+                        {"role": role, "content": ". ".join(sentences) + "."}
+                    )
+                    continue
+                else:
+                    mutated_messages.append({"role": role, "content": content})
+
+        return chat_prompt.ChatPrompt(messages=mutated_messages)
+
+    def _word_level_mutation_prompt(
+        self, prompt: chat_prompt.ChatPrompt
+    ) -> chat_prompt.ChatPrompt:
+        mutated_messages: list[dict[str, str]] = []
+        for message in prompt.get_messages():
+            mutated_messages.append(
+                {
+                    "role": message["role"],
+                    "content": self._word_level_mutation(message["content"]),
+                }
+            )
+        return chat_prompt.ChatPrompt(messages=mutated_messages)
+
+    def _word_level_mutation(self, msg_content: str) -> str:
+        """Perform word-level mutation."""
+        words = msg_content.split()
+        if len(words) <= 1:
+            return msg_content
+
+        mutation_type = random.random()
+        if mutation_type < 0.3:
+            idx = random.randint(0, len(words) - 1)
+            words[idx] = self._get_synonym(words[idx])
+        elif mutation_type < 0.6:
+            if len(words) > 2:
+                i, j = random.sample(range(len(words)), 2)
+                words[i], words[j] = words[j], words[i]
+        else:
+            idx = random.randint(0, len(words) - 1)
+            words[idx] = self._modify_phrase(words[idx])
+
+        return " ".join(words)
+
+    def _get_synonym(self, word: str) -> str:
+        """Get a synonym for a word using LLM."""
+        try:
+            response = self._call_model(
+                messages=[
+                    {"role": "system", "content": evo_prompts.synonyms_system_prompt()},
+                    {
+                        "role": "user",
+                        "content": (
+                            f"Give me a single synonym for the word '{word}'. Return only the synonym, nothing else."
+                        ),
+                    },
+                ],
+                is_reasoning=True,
+            )
+            return response.strip()
+        except Exception as e:
+            logger.warning(f"Error getting synonym for '{word}': {e}")
+            return word
+
+    def _modify_phrase(self, phrase: str) -> str:
+        """Modify a phrase while preserving meaning using LLM."""
+        try:
+            response = self._call_model(
+                messages=[
+                    {"role": "system", "content": evo_prompts.rephrase_system_prompt()},
+                    {
+                        "role": "user",
+                        "content": (
+                            f"Modify this phrase while keeping the same meaning: '{phrase}'. Return only the modified phrase, nothing else."
+                        ),
+                    },
+                ],
+                is_reasoning=True,
+            )
+            return response.strip()
+        except Exception as e:
+            logger.warning(f"Error modifying phrase '{phrase}': {e}")
+            return phrase
+
+    def _radical_innovation_mutation(
+        self, prompt: chat_prompt.ChatPrompt, initial_prompt: chat_prompt.ChatPrompt
+    ) -> chat_prompt.ChatPrompt:
+        """Attempts to generate a significantly improved and potentially very different prompt using an LLM."""
+        logger.debug(
+            f"Attempting radical innovation for prompt: {json.dumps(prompt.get_messages())[:70]}..."
+        )
+        task_desc_for_llm = self._get_task_description_for_llm(initial_prompt)
+        current_output_style_guidance = self.output_style_guidance
+
+        user_prompt_for_radical_innovation = evo_prompts.radical_innovation_user_prompt(
+            task_desc_for_llm, current_output_style_guidance, prompt.get_messages()
+        )
+        try:
+            new_prompt_str = self._call_model(
+                messages=[
+                    {
+                        "role": "system",
+                        "content": evo_prompts.radical_innovation_system_prompt(
+                            current_output_style_guidance
+                        ),
+                    },
+                    {"role": "user", "content": user_prompt_for_radical_innovation},
+                ],
+                is_reasoning=True,
+            )
+            logger.info(
+                f"Radical innovation LLM result (truncated): {new_prompt_str[:200]}"
+            )
+            try:
+                new_messages = utils.json_to_dict(new_prompt_str)
+            except Exception as parse_exc:
+                logger.warning(
+                    f"Failed to parse LLM output in radical innovation mutation for prompt '{json.dumps(prompt.get_messages())[:50]}...'. Output: {new_prompt_str[:200]}. Error: {parse_exc}. Returning original."
+                )
+                return prompt
+            return chat_prompt.ChatPrompt(messages=new_messages)
+        except Exception as e:
+            logger.warning(
+                f"Radical innovation mutation failed for prompt '{json.dumps(prompt.get_messages())[:50]}...': {e}. Returning original."
+            )
+            return prompt

opik_optimizer/evolutionary_optimizer/population_ops.py

@@ -0,0 +1,223 @@
+from typing import Any, TYPE_CHECKING
+
+import json
+import logging
+
+from deap import tools
+from deap import creator as _creator
+
+from . import prompts as evo_prompts
+from . import reporting
+from ..optimization_config import chat_prompt
+from .. import utils
+
+
+logger = logging.getLogger(__name__)
+creator = _creator
+
+
+class PopulationOps:
+    if TYPE_CHECKING:
+        _get_task_description_for_llm: Any
+        output_style_guidance: str
+        _call_model: Any
+        toolbox: Any
+    # Hints for mixin attributes provided by the primary optimizer class
+    _gens_since_pop_improvement: int
+    _best_primary_score_history: list[float]
+    DEFAULT_RESTART_THRESHOLD: float
+    DEFAULT_RESTART_GENERATIONS: int
+    enable_moo: bool
+    elitism_size: int
+    population_size: int
+    verbose: int
+
+    def _initialize_population(
+        self, prompt: chat_prompt.ChatPrompt
+    ) -> list[chat_prompt.ChatPrompt]:
+        """Initialize the population with diverse variations of the initial prompt,
+        including some 'fresh start' prompts based purely on task description.
+        All generated prompts should aim to elicit answers matching self.output_style_guidance.
+        """
+        with reporting.initializing_population(verbose=self.verbose) as init_pop_report:
+            init_pop_report.start(self.population_size)
+
+            population = [prompt]
+            if self.population_size <= 1:
+                return population
+
+            num_to_generate_total = self.population_size - 1
+            num_fresh_starts = max(1, int(num_to_generate_total * 0.2))
+            num_variations_on_initial = num_to_generate_total - num_fresh_starts
+
+            task_desc_for_llm = self._get_task_description_for_llm(prompt)
+            current_output_style_guidance = self.output_style_guidance
+
+            # Fresh starts
+            if num_fresh_starts > 0:
+                init_pop_report.start_fresh_prompts(num_fresh_starts)
+                fresh_start_user_prompt = evo_prompts.fresh_start_user_prompt(
+                    task_desc_for_llm, current_output_style_guidance, num_fresh_starts
+                )
+                try:
+                    response_content = self._call_model(
+                        messages=[
+                            {
+                                "role": "system",
+                                "content": evo_prompts.fresh_start_system_prompt(
+                                    current_output_style_guidance
+                                ),
+                            },
+                            {"role": "user", "content": fresh_start_user_prompt},
+                        ],
+                        is_reasoning=True,
+                    )
+
+                    logger.debug(
+                        f"Raw LLM response for fresh start prompts: {response_content}"
+                    )
+
+                    fresh_prompts = utils.json_to_dict(response_content)
+                    if isinstance(fresh_prompts, list):
+                        if all(isinstance(p, dict) for p in fresh_prompts) and all(
+                            p.get("role") is not None for p in fresh_prompts
+                        ):
+                            population.append(
+                                chat_prompt.ChatPrompt(messages=fresh_prompts)
+                            )
+                            init_pop_report.success_fresh_prompts(1)
+                        elif all(isinstance(p, list) for p in fresh_prompts):
+                            population.extend(
+                                [
+                                    chat_prompt.ChatPrompt(messages=p)
+                                    for p in fresh_prompts[:num_fresh_starts]
+                                ]
+                            )
+                            init_pop_report.success_fresh_prompts(
+                                len(fresh_prompts[:num_fresh_starts])
+                            )
+                        else:
+                            init_pop_report.failed_fresh_prompts(
+                                num_fresh_starts,
+                                f"LLM response for fresh starts was not a valid list of strings or was empty: {response_content}. Skipping fresh start prompts.",
+                            )
+                except json.JSONDecodeError as e_json:
+                    init_pop_report.failed_fresh_prompts(
+                        num_fresh_starts,
+                        f"JSONDecodeError generating fresh start prompts: {e_json}. LLM response: '{response_content}'. Skipping fresh start prompts.",
+                    )
+                except Exception as e:
+                    init_pop_report.failed_fresh_prompts(
+                        num_fresh_starts,
+                        f"Error generating fresh start prompts: {e}. Skipping fresh start prompts.",
+                    )
+
+            # Variations on the initial prompt
+            if num_variations_on_initial > 0:
+                init_pop_report.start_variations(num_variations_on_initial)
+                user_prompt_for_variation = evo_prompts.variation_user_prompt(
+                    prompt.get_messages(),
+                    task_desc_for_llm,
+                    current_output_style_guidance,
+                    num_variations_on_initial,
+                )
+                try:
+                    response_content_variations = self._call_model(
+                        messages=[
+                            {
+                                "role": "system",
+                                "content": evo_prompts.variation_system_prompt(
+                                    current_output_style_guidance
+                                ),
+                            },
+                            {"role": "user", "content": user_prompt_for_variation},
+                        ],
+                        is_reasoning=True,
+                    )
+                    logger.debug(
+                        f"Raw response for population variations: {response_content_variations}"
+                    )
+                    json_response_variations = json.loads(response_content_variations)
+                    generated_prompts_variations = [
+                        p["prompt"]
+                        for p in json_response_variations.get("prompts", [])
+                        if isinstance(p, dict) and "prompt" in p
+                    ]
+
+                    if generated_prompts_variations:
+                        init_pop_report.success_variations(
+                            len(
+                                generated_prompts_variations[:num_variations_on_initial]
+                            )
+                        )
+                        population.extend(
+                            [
+                                chat_prompt.ChatPrompt(messages=p)
+                                for p in generated_prompts_variations[
+                                    :num_variations_on_initial
+                                ]
+                            ]
+                        )
+                    else:
+                        init_pop_report.failed_variations(
+                            num_variations_on_initial,
+                            "Could not parse 'prompts' list for variations. Skipping variations.",
+                        )
+                except Exception as e:
+                    init_pop_report.failed_variations(
+                        num_variations_on_initial,
+                        f"Error calling LLM for initial population variations: {e}",
+                    )
+
+            # Ensure population is of the required size using unique prompts
+            final_population_set: set[str] = set()
+            final_population_list: list[chat_prompt.ChatPrompt] = []
+            for p in population:
+                if json.dumps(p.get_messages()) not in final_population_set:
+                    final_population_set.add(json.dumps(p.get_messages()))
+                    final_population_list.append(p)
+
+            init_pop_report.end(final_population_list)
+            return final_population_list[: self.population_size]
+
+    def _should_restart_population(self, curr_best: float) -> bool:
+        """Update internal counters and decide if we should trigger a population restart."""
+        if self._best_primary_score_history:
+            threshold = self._best_primary_score_history[-1] * (
+                1 + self.DEFAULT_RESTART_THRESHOLD
+            )
+            if curr_best < threshold:
+                self._gens_since_pop_improvement += 1  # type: ignore[attr-defined]
+            else:
+                self._gens_since_pop_improvement = 0  # type: ignore[attr-defined]
+        self._best_primary_score_history.append(curr_best)
+        return self._gens_since_pop_improvement >= self.DEFAULT_RESTART_GENERATIONS  # type: ignore[attr-defined]
+
+    def _restart_population(
+        self,
+        hof: tools.HallOfFame,
+        population: list[Any],
+        best_prompt_so_far: chat_prompt.ChatPrompt,
+    ) -> list[Any]:
+        """Return a fresh, evaluated population seeded by elites."""
+        if self.enable_moo:
+            elites = list(hof)
+        else:
+            elites = tools.selBest(population, self.elitism_size)
+
+        seed_prompt = (
+            chat_prompt.ChatPrompt(
+                messages=max(elites, key=lambda x: x.fitness.values[0])
+            )
+            if elites
+            else best_prompt_so_far
+        )
+
+        prompt_variants = self._initialize_population(seed_prompt)
+        new_pop = [creator.Individual(p.get_messages()) for p in prompt_variants]
+
+        for ind, fit in zip(new_pop, map(self.toolbox.evaluate, new_pop)):
+            ind.fitness.values = fit
+
+        self._gens_since_pop_improvement = 0  # type: ignore[attr-defined]
+        return new_pop