AgentOptimizer

agential/agents/react/functional.py

@@ -33,7 +33,7 @@ def _build_agent_prompt(
 
     Returns:
         str: A formatted prompt template ready for use.
-    """
+    """          
     prompt = prompt.format(
         question=question,
         scratchpad=scratchpad,

agential/training/agent_optimizer/__init__.py

@@ -0,0 +1 @@
+"""Prompt Optimizer Agent."""

agential/training/agent_optimizer/agent.py

@@ -0,0 +1,251 @@
+"""ReAct Agent.
+
+Original Paper: https://arxiv.org/abs/2210.03629
+Paper Repository: https://github.com/ysymyth/ReAct
+"""
+
+from typing import Any, Dict, Optional
+
+from agential.agents.base.agent import BaseAgent
+from agential.training.agent_optimizer.output import PromptOptimizerOutput
+from agential.training.agent_optimizer.prompts import (
+    PROMPT_OPTIMIZER_INSTRUCTION_AMBIGNQ,
+    PROMPT_OPTIMIZER_INSTRUCTION_FEVER,
+    PROMPT_OPTIMIZER_INSTRUCTION_GSM8K,
+    PROMPT_OPTIMIZER_INSTRUCTION_HOTPOTQA,
+    PROMPT_OPTIMIZER_INSTRUCTION_HUMANEVAL,
+    PROMPT_OPTIMIZER_INSTRUCTION_MBPP,
+    PROMPT_OPTIMIZER_INSTRUCTION_SVAMP,
+    PROMPT_OPTIMIZER_INSTRUCTION_TABMWP,
+    PROMPT_OPTIMIZER_INSTRUCTION_TRIVIAQA,
+)
+from agential.training.agent_optimizer.strategies.base import PromptOptimizerBaseStrategy
+from agential.training.agent_optimizer.strategies.code import PromptOptimizerHEvalStrategy, PromptOptimizerMBPPStrategy
+from agential.training.agent_optimizer.strategies.math import (
+    PromptOptimizerGSM8KStrategy,
+    PromptOptimizerSVAMPStrategy,
+    PromptOptimizerTabMWPStrategy,
+)
+from agential.training.agent_optimizer.strategies.qa import (
+    PromptOptimizerAmbigNQStrategy,
+    PromptOptimizerFEVERStrategy,
+    PromptOptimizerHotQAStrategy,
+    PromptOptimizerTriviaQAStrategy,
+)
+from agential.constants import BENCHMARK_FEWSHOTS, Benchmarks, FewShotType
+from agential.core.llm import BaseLLM
+
+PROMPT_OPTIMIZER_BENCHMARK_FEWSHOTS = {
+    Benchmarks.HOTPOTQA: [FewShotType.REACT],
+    Benchmarks.FEVER: [FewShotType.REACT],
+    Benchmarks.TRIVIAQA: [FewShotType.REACT],
+    Benchmarks.AMBIGNQ: [FewShotType.REACT],
+    Benchmarks.GSM8K: [FewShotType.REACT],
+    Benchmarks.SVAMP: [FewShotType.REACT],
+    Benchmarks.TABMWP: [FewShotType.REACT],
+    Benchmarks.HUMANEVAL: [FewShotType.REACT],
+    Benchmarks.MBPP: [FewShotType.REACT],
+}
+
+REACT_PROMPTS = {
+    Benchmarks.HOTPOTQA: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_HOTPOTQA,
+    },
+    Benchmarks.FEVER: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_FEVER,
+    },
+    Benchmarks.TRIVIAQA: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_TRIVIAQA,
+    },
+    Benchmarks.AMBIGNQ: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_AMBIGNQ,
+    },
+    Benchmarks.GSM8K: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_GSM8K,
+    },
+    Benchmarks.SVAMP: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_SVAMP,
+    },
+    Benchmarks.TABMWP: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_TABMWP,
+    },
+    Benchmarks.HUMANEVAL: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_HUMANEVAL,
+    },
+    Benchmarks.MBPP: {
+        "prompt": PROMPT_OPTIMIZER_INSTRUCTION_MBPP,
+    },
+}
+REACT_FEWSHOTS: Dict[str, Dict] = {
+    Benchmarks.HOTPOTQA: {},
+    Benchmarks.FEVER: {},
+    Benchmarks.TRIVIAQA: {},
+    Benchmarks.AMBIGNQ: {},
+    Benchmarks.GSM8K: {},
+    Benchmarks.SVAMP: {},
+    Benchmarks.TABMWP: {},
+    Benchmarks.HUMANEVAL: {},
+    Benchmarks.MBPP: {},
+}
+REACT_STRATEGIES = {
+    Benchmarks.HOTPOTQA: PromptOptimizerHotQAStrategy,
+    Benchmarks.FEVER: PromptOptimizerFEVERStrategy,
+    Benchmarks.TRIVIAQA: PromptOptimizerTriviaQAStrategy,
+    Benchmarks.AMBIGNQ: PromptOptimizerAmbigNQStrategy,
+    Benchmarks.GSM8K: PromptOptimizerGSM8KStrategy,
+    Benchmarks.SVAMP: PromptOptimizerSVAMPStrategy,
+    Benchmarks.TABMWP: PromptOptimizerTabMWPStrategy,
+    Benchmarks.HUMANEVAL: PromptOptimizerHEvalStrategy,
+    Benchmarks.MBPP: PromptOptimizerMBPPStrategy,
+}
+
+
+class PromptOptimizer(BaseAgent):
+    """PromptOptimizer class for optimizing the agent's performance."""
+
+    def __init__(
+        self,
+        llm: BaseLLM,
+        benchmark: str,
+        testing: bool = False,
+        max_steps: int = 3,
+        optimizer_model: Optional[str] = "gpt-4-1106-preview",
+        **strategy_kwargs: Any,
+    ) -> None:
+        """Initialization."""
+        super().__init__(llm=llm, benchmark=benchmark, testing=testing)
+
+
+class PromptOptimizer(BaseAgent):
+    """PromptOptimizer agent.
+
+    Attributes:
+        llm (BaseLLM): An instance of a language model used for generating initial answers
+            and critiques.
+        benchmark (str): The benchmark.
+        testing (bool, optional): Whether to run in testing mode. Defaults to False.
+        **strategy_kwargs (Any): Additional strategy-specific arguments.
+    """
+
+    def __init__(
+        self,
+        llm: BaseLLM,
+        benchmark: str,
+        testing: bool = False,
+        **strategy_kwargs: Any,
+    ) -> None:
+        """Initialization."""
+        super().__init__(llm=llm, benchmark=benchmark, testing=testing)
+
+        self.strategy = PromptOptimizer.get_strategy(
+            benchmark=self.benchmark,
+            llm=self.llm,
+            testing=self.testing,
+            **strategy_kwargs,
+        )
+
+    @staticmethod
+    def get_fewshots(
+        benchmark: str, fewshot_type: str, **kwargs: Any
+    ) -> Dict[str, str]:
+        """Retrieve few-shot examples based on the benchmark.
+
+        Args:
+            benchmark (str): The benchmark name.
+            fewshot_type (str): The benchmark few-shot type.
+            **kwargs (Any): Additional arguments.
+
+        Returns:
+            Dict[str, str]: A dictionary of few-shot examples.
+        """
+        if benchmark not in REACT_FEWSHOTS:
+            raise ValueError(f"Benchmark '{benchmark}' few-shots not found for ReAct.")
+
+        if fewshot_type not in PROMPT_OPTIMIZER_BENCHMARK_FEWSHOTS[benchmark]:
+            raise ValueError(
+                f"Benchmark '{benchmark}' few-shot type not supported for ReAct."
+            )
+
+        benchmark_fewshots = BENCHMARK_FEWSHOTS[benchmark][fewshot_type]
+
+        return {"examples": benchmark_fewshots}
+
+    @staticmethod
+    def get_prompts(benchmark: str, **kwargs: Any) -> Dict[str, str]:
+        """Retrieve the prompt instruction based on the benchmark.
+
+        Args:
+            benchmark (str): The benchmark name.
+            **kwargs (Any): Additional arguments.
+
+        Returns:
+            Dict[str, str]: A dictionary of prompt instructions.
+        """
+        if benchmark not in REACT_PROMPTS:
+            raise ValueError(f"Benchmark '{benchmark}' prompt not found for ReAct.")
+
+        return REACT_PROMPTS[benchmark]
+
+    @staticmethod
+    def get_strategy(benchmark: str, **kwargs: Any) -> PromptOptimizerBaseStrategy:
+        """Returns an instance of the appropriate ReAct strategy based on the provided benchmark.
+
+        Args:
+            benchmark (str): The benchmark name.
+            **kwargs (Any): Additional keyword arguments to pass to
+                the strategy's constructor.
+
+        Returns:
+            ReActBaseStrategy: An instance of the appropriate ReAct strategy.
+        """
+        if benchmark not in REACT_STRATEGIES:
+            raise ValueError(f"Unsupported benchmark: {benchmark} for agent ReAct")
+
+        strategy = REACT_STRATEGIES[benchmark]
+        return strategy(**kwargs)
+
+    def generate(
+        self,
+        question: str,
+        examples: str = "",
+        prompt: str = "",
+        additional_keys: Dict[str, str] = {},
+        fewshot_type: str = "",
+        reset: bool = True,
+    ) -> PromptOptimizerOutput:
+        """Processes a given question through ReAct.
+
+        Iteratively applies the think-act-observe cycle to generate an answer for the question.
+        The process continues until the operation is halted based on certain conditions.
+
+        Args:
+            question (str): The question to be processed.
+            examples (str, optional): Fewshot examples. Defaults to "".
+            prompt (str, optional): Prompt template string. Defaults to "".
+            additional_keys (Dict[str, str]): Additional keys to format the prompt. Defaults to {}.
+            fewshot_type (str): The type of few-shot examples to use. Defaults to "".
+            reset (bool, optional): Whether to reset the internal state before processing. Defaults to True.
+
+        Returns:
+            ReActOutput: The list of accumulated output from the ReAct process,
+                each ReActOutput consists of a thought, action type/query, observation, answer, and external tool info.
+        """
+        if not prompt or not examples:
+            if not fewshot_type:
+                fewshot_type = PROMPT_OPTIMIZER_BENCHMARK_FEWSHOTS[self.benchmark][0]
+            fewshots = PromptOptimizer.get_fewshots(
+                benchmark=self.benchmark, fewshot_type=fewshot_type
+            )
+            prompts = PromptOptimizer.get_prompts(benchmark=self.benchmark)
+            examples = fewshots["examples"]
+            prompt = prompts["prompt"]
+
+        out = self.strategy.generate(
+            question=question,
+            examples=examples,
+            prompt=prompt,
+            additional_keys=additional_keys,
+            reset=reset,
+        )
+
+        return out