chatterer 0.1.5__py3-none-any.whl → 0.1.6__py3-none-any.whl

chatterer/strategies/atom_of_thoughts.py

@@ -1,32 +1,25 @@
-# original source: https://github.com/qixucen/atom
-
 from __future__ import annotations
 
 import asyncio
 import logging
-from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
 from enum import StrEnum
-from typing import LiteralString, Optional, Type, TypeVar
+from typing import Optional, Type, TypeVar
 
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage
 from pydantic import BaseModel, Field, ValidationError
 
+# Import your Chatterer interface (do not remove)
 from ..language_model import Chatterer, LanguageModelInput
 from .base import BaseStrategy
 
-logger = logging.getLogger(__name__)
-
-# ------------------------- 0) Enums and Basic Models -------------------------
-
+# ---------------------------------------------------------------------------------
+# 0) Enums and Basic Models
+# ---------------------------------------------------------------------------------
 
-class Domain(StrEnum):
-    """Defines the domain of a question for specialized handling."""
-
-    GENERAL = "general"
-    MATH = "math"
-    CODING = "coding"
-    PHILOSOPHY = "philosophy"
-    MULTIHOP = "multihop"
+QA_TEMPLATE = "Q: {question}\nA: {answer}"
+MAX_DEPTH_REACHED = "Max depth reached in recursive decomposition."
+UNKNOWN = "Unknown"
 
 
 class SubQuestionNode(BaseModel):
@@ -45,13 +38,6 @@ class RecursiveDecomposeResponse(BaseModel):
     sub_questions: list[SubQuestionNode] = Field(description="Root-level sub-questions.")
 
 
-class DirectResponse(BaseModel):
-    """A direct response to a question."""
-
-    thought: str = Field(description="Short reasoning.")
-    answer: str = Field(description="Direct final answer.")
-
-
 class ContractQuestionResponse(BaseModel):
     """The result of contracting (simplifying) a question."""
 
@@ -66,359 +52,507 @@ class EnsembleResponse(BaseModel):
     answer: str = Field(description="Best final answer after ensemble.")
     confidence: float = Field(description="Confidence score in [0, 1].")
 
-    def model_post_init(self, __context) -> None:
-        # Clamp confidence to [0, 1]
+    def model_post_init(self, __context: object) -> None:
         self.confidence = max(0.0, min(1.0, self.confidence))
 
 
 class LabelResponse(BaseModel):
-    """A response used to refine sub-question dependencies and structure."""
+    """Used to refine and reorder the sub-questions with corrected dependencies."""
 
     thought: str = Field(description="Explanation or reasoning about labeling.")
     sub_questions: list[SubQuestionNode] = Field(
         description="Refined list of sub-questions with corrected dependencies."
     )
-    # Some tasks also keep the final answer, but we focus on sub-questions.
 
 
-# --------------- 1) Prompter Classes with multi-hop context ---------------
+class CritiqueResponse(BaseModel):
+    """A response used for LLM to self-critique or question its own correctness."""
 
+    thought: str = Field(description="Critical reflection on correctness.")
+    self_assessment: float = Field(description="Self-assessed confidence in the approach/answer. A float in [0,1].")
 
-class BaseAoTPrompter(ABC):
-    """
-    Abstract base prompter that defines the required prompt methods.
-    """
 
-    @abstractmethod
-    def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str: ...
+# ---------------------------------------------------------------------------------
+# [NEW] Additional classes to incorporate a separate sub-question devil's advocate
+# ---------------------------------------------------------------------------------
 
-    @abstractmethod
-    def direct_prompt(self, question: str, context: Optional[str] = None) -> str: ...
 
-    @abstractmethod
-    def label_prompt(
-        self, question: str, decompose_response: RecursiveDecomposeResponse, context: Optional[str] = None
-    ) -> str:
-        """
-        Prompt used to 're-check' the sub-questions for correctness and dependencies.
-        """
+class DevilsAdvocateResponse(BaseModel):
+    """
+    A response for a 'devil's advocate' pass.
+    We consider an alternative viewpoint or contradictory answer.
+    """
 
-    @abstractmethod
-    def contract_prompt(self, question: str, sub_answers: str, context: Optional[str] = None) -> str: ...
+    thought: str = Field(description="Reasoning behind the contradictory viewpoint.")
+    final_answer: str = Field(description="Alternative or conflicting answer to challenge the main one.")
+    sub_questions: list[SubQuestionNode] = Field(
+        description="Any additional sub-questions from the contrarian perspective."
+    )
 
-    @abstractmethod
-    def ensemble_prompt(
-        self,
-        original_question: str,
-        direct_answer: str,
-        decompose_answer: str,
-        contracted_direct_answer: str,
-        context: Optional[str] = None,
-    ) -> str: ...
 
+# ---------------------------------------------------------------------------------
+# 1) Prompter Classes with Multi-Hop + Devil's Advocate
+# ---------------------------------------------------------------------------------
 
-class GeneralAoTPrompter(BaseAoTPrompter):
-    """
-    Generic prompter for non-specialized or 'general' queries.
-    """
+
+class AoTPrompter:
+    """Generic base prompter that defines the required prompt methods."""
 
     def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str:
-        sub_ans_str = f"\nSub-question answers:\n{sub_answers}" if sub_answers else ""
-        context_str = f"\nCONTEXT:\n{context}" if context else ""
-        return (
-            "You are a highly analytical assistant skilled in breaking down complex problems.\n"
-            "Decompose the question into sub-questions recursively.\n\n"
-            "REQUIREMENTS:\n"
-            "1. Return valid JSON:\n"
-            "   {\n"
-            '     "thought": "...",\n'
-            '     "final_answer": "...",\n'
-            '     "sub_questions": [{"question": "...", "answer": null, "depend": []}, ...]\n'
-            "   }\n"
-            "2. 'thought': Provide detailed reasoning.\n"
-            "3. 'final_answer': Integrate sub-answers if any.\n"
-            "4. 'sub_questions': Key sub-questions with potential dependencies.\n\n"
-            f"QUESTION:\n{question}{sub_ans_str}{context_str}"
+        self, messages: list[BaseMessage], question: str, sub_answers: list[tuple[str, str]]
+    ) -> list[BaseMessage]:
+        """
+        Prompt for main decomposition.
+        Encourages step-by-step reasoning and listing sub-questions as JSON.
+        """
+        decompose_instructions = (
+            "First, restate the main question.\n"
+            "Decide if sub-questions are needed. If so, list them.\n"
+            "In the 'thought' field, show your chain-of-thought.\n"
+            "Return valid JSON:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "final_answer": "...",\n'
+            '  "sub_questions": [\n'
+            '    {"question": "...", "answer": null, "depend": []},\n'
+            "    ...\n"
+            "  ]\n"
+            "}\n"
         )
 
-    def direct_prompt(self, question: str, context: Optional[str] = None) -> str:
-        context_str = f"\nCONTEXT:\n{context}" if context else ""
-        return (
-            "You are a concise and insightful assistant.\n"
-            "Provide a direct answer with a short reasoning.\n\n"
-            "REQUIREMENTS:\n"
-            "1. Return valid JSON:\n"
-            "   {'thought': '...', 'answer': '...'}\n"
-            "2. 'thought': Offer a brief reasoning.\n"
-            "3. 'answer': Deliver a precise solution.\n\n"
-            f"QUESTION:\n{question}{context_str}"
-        )
+        content_sub_answers = "\n".join(f"Sub-answer so far: Q={q}, A={a}" for q, a in sub_answers)
+        return messages + [
+            HumanMessage(content=f"Main question:\n{question}"),
+            AIMessage(content=content_sub_answers),
+            AIMessage(content=decompose_instructions),
+        ]
 
     def label_prompt(
-        self, question: str, decompose_response: RecursiveDecomposeResponse, context: Optional[str] = None
-    ) -> str:
-        context_str = f"\nCONTEXT:\n{context}" if context else ""
-        return (
-            "You have a set of sub-questions from a decomposition process.\n"
-            "We want to correct or refine the dependencies between sub-questions.\n\n"
-            "REQUIREMENTS:\n"
-            "1. Return valid JSON:\n"
-            "   {\n"
-            '     "thought": "...",\n'
-            '     "sub_questions": [\n'
-            '         {"question":"...", "answer":"...", "depend":[...]},\n'
-            "         ...\n"
-            "     ]\n"
-            "   }\n"
-            "2. 'thought': Provide reasoning about any changes.\n"
-            "3. 'sub_questions': Possibly updated sub-questions with correct 'depend' lists.\n\n"
-            f"ORIGINAL QUESTION:\n{question}\n"
-            f"CURRENT DECOMPOSITION:\n{decompose_response.model_dump_json(indent=2)}"
-            f"{context_str}"
+        self, messages: list[BaseMessage], question: str, decompose_response: RecursiveDecomposeResponse
+    ) -> list[BaseMessage]:
+        """
+        Prompt for refining the sub-questions and dependencies.
+        """
+        label_instructions = (
+            "Review each sub-question to ensure correctness and proper ordering.\n"
+            "Return valid JSON in the form:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "sub_questions": [\n'
+            '    {"question": "...", "answer": "...", "depend": [...]},\n'
+            "    ...\n"
+            "  ]\n"
+            "}\n"
         )
+        return messages + [
+            AIMessage(content=f"Question: {question}"),
+            AIMessage(content=f"Current sub-questions:\n{decompose_response.sub_questions}"),
+            AIMessage(content=label_instructions),
+        ]
 
-    def contract_prompt(self, question: str, sub_answers: str, context: Optional[str] = None) -> str:
-        context_str = f"\nCONTEXT:\n{context}" if context else ""
-        return (
-            "You are tasked with compressing or simplifying a complex question into a single self-contained one.\n\n"
-            "REQUIREMENTS:\n"
-            "1. Return valid JSON:\n"
-            "   {'thought': '...', 'question': '...'}\n"
-            "2. 'thought': Explain your simplification.\n"
-            "3. 'question': The streamlined question.\n\n"
-            f"ORIGINAL QUESTION:\n{question}\n"
-            f"SUB-ANSWERS:\n{sub_answers}"
-            f"{context_str}"
+    def contract_prompt(self, messages: list[BaseMessage], sub_answers: list[tuple[str, str]]) -> list[BaseMessage]:
+        """
+        Prompt for merging sub-answers into one self-contained question.
+        """
+        contract_instructions = (
+            "Please merge sub-answers into a single short question that is fully self-contained.\n"
+            "In 'thought', show how you unify the information.\n"
+            "Then produce JSON:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "question": "a short but self-contained question"\n'
+            "}\n"
         )
-
-    def ensemble_prompt(
-        self,
-        original_question: str,
-        direct_answer: str,
-        decompose_answer: str,
-        contracted_direct_answer: str,
-        context: Optional[str] = None,
-    ) -> str:
-        context_str = f"\nCONTEXT:\n{context}" if context else ""
-        return (
-            "You are an expert at synthesizing multiple candidate answers.\n"
-            "Consider the following candidates:\n"
-            f"1) Direct: {direct_answer}\n"
-            f"2) Decomposition-based: {decompose_answer}\n"
-            f"3) Contracted Direct: {contracted_direct_answer}\n\n"
-            "REQUIREMENTS:\n"
-            "1. Return valid JSON:\n"
-            "   {'thought': '...', 'answer': '...', 'confidence': <float in [0,1]>}\n"
-            "2. 'thought': Summarize how you decided.\n"
-            "3. 'answer': Final best answer.\n"
-            "4. 'confidence': Confidence score in [0,1].\n\n"
-            f"ORIGINAL QUESTION:\n{original_question}"
-            f"{context_str}"
+        sub_q_content = "\n".join(f"Q: {q}\nA: {a}" for q, a in sub_answers)
+        return messages + [
+            AIMessage(content="We have these sub-questions and answers:"),
+            AIMessage(content=sub_q_content),
+            AIMessage(content=contract_instructions),
+        ]
+
+    def contract_direct_prompt(self, messages: list[BaseMessage], contracted_question: str) -> list[BaseMessage]:
+        """
+        Prompt for directly answering the contracted question thoroughly.
+        """
+        direct_instructions = (
+            "Answer the simplified question thoroughly. Show your chain-of-thought in 'thought'.\n"
+            "Return JSON:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "final_answer": "..."\n'
+            "}\n"
         )
+        return messages + [
+            HumanMessage(content=f"Simplified question: {contracted_question}"),
+            AIMessage(content=direct_instructions),
+        ]
 
-
-class MathAoTPrompter(GeneralAoTPrompter):
-    """
-    Specialized prompter for math questions; includes domain-specific hints.
-    """
-
-    def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str:
-        base = super().recursive_decompose_prompt(question, sub_answers, context)
-        return base + "\nFocus on mathematical rigor and step-by-step derivations.\n"
-
-    def direct_prompt(self, question: str, context: Optional[str] = None) -> str:
-        base = super().direct_prompt(question, context)
-        return base + "\nEnsure mathematical correctness in your reasoning.\n"
-
-    # label_prompt, contract_prompt, ensemble_prompt can also be overridden similarly if needed.
-
-
-class CodingAoTPrompter(GeneralAoTPrompter):
-    """
-    Specialized prompter for coding/algorithmic queries.
-    """
-
-    def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str:
-        base = super().recursive_decompose_prompt(question, sub_answers, context)
-        return base + "\nBreak down into programming concepts or implementation steps.\n"
-
-
-class PhilosophyAoTPrompter(GeneralAoTPrompter):
-    """
-    Specialized prompter for philosophical discussions.
-    """
-
-    def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str:
-        base = super().recursive_decompose_prompt(question, sub_answers, context)
-        return base + "\nConsider key philosophical theories and arguments.\n"
-
-
-class MultiHopAoTPrompter(GeneralAoTPrompter):
-    """
-    Specialized prompter for multi-hop Q&A with explicit context usage.
-    """
-
-    def recursive_decompose_prompt(
-        self, question: str, sub_answers: Optional[str] = None, context: Optional[str] = None
-    ) -> str:
-        base = super().recursive_decompose_prompt(question, sub_answers, context)
-        return (
-            base + "\nTreat this as a multi-hop question. Use the provided context carefully.\n"
-            "Extract partial evidence from the context for each sub-question.\n"
+    def critique_prompt(self, messages: list[BaseMessage], thought: str, answer: str) -> list[BaseMessage]:
+        """
+        Prompt for self-critique.
+        """
+        critique_instructions = (
+            "Critique your own approach. Identify possible errors or leaps.\n"
+            "Return JSON:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "self_assessment": <float in [0,1]>\n'
+            "}\n"
         )
+        return messages + [
+            AIMessage(content=f"Your previous THOUGHT:\n{thought}"),
+            AIMessage(content=f"Your previous ANSWER:\n{answer}"),
+            AIMessage(content=critique_instructions),
+        ]
 
-    def direct_prompt(self, question: str, context: Optional[str] = None) -> str:
-        base = super().direct_prompt(question, context)
-        return base + "\nFor multi-hop, ensure each piece of reasoning uses only the relevant parts of the context.\n"
-
-
-# ----------------- 2) The AoTPipeline class with label + score + multi-hop -----------------
-
-T = TypeVar("T", bound=BaseModel)
+    def ensemble_prompt(
+        self, messages: list[BaseMessage], possible_thought_and_answers: list[tuple[str, str]]
+    ) -> list[BaseMessage]:
+        """
+        Show multiple candidate solutions and pick the best final answer with confidence.
+        """
+        instructions = (
+            "You have multiple candidate solutions. Compare carefully and pick the best.\n"
+            "Return JSON:\n"
+            "{\n"
+            '  "thought": "why you chose this final answer",\n'
+            '  "answer": "the best consolidated answer",\n'
+            '  "confidence": 0.0 ~ 1.0\n'
+            "}\n"
+        )
+        reasonings: list[BaseMessage] = []
+        for idx, (thought, ans) in enumerate(possible_thought_and_answers):
+            reasonings.append(AIMessage(content=f"[Candidate {idx}] Thought:\n{thought}\nAnswer:\n{ans}\n---"))
+        return messages + reasonings + [AIMessage(content=instructions)]
+
+    def devils_advocate_prompt(
+        self, messages: list[BaseMessage], question: str, existing_answer: str
+    ) -> list[BaseMessage]:
+        """
+        Prompt for a devil's advocate approach to contradict or provide an alternative viewpoint.
+        """
+        instructions = (
+            "Act as a devil's advocate. Suppose the existing answer is incomplete or incorrect.\n"
+            "Challenge it, find alternative ways or details. Provide a new 'final_answer' (even if contradictory).\n"
+            "Return JSON in the same shape as RecursiveDecomposeResponse OR a dedicated structure.\n"
+            "But here, let's keep it in a new dedicated structure:\n"
+            "{\n"
+            '  "thought": "...",\n'
+            '  "final_answer": "...",\n'
+            '  "sub_questions": [\n'
+            '    {"question": "...", "answer": null, "depend": []},\n'
+            "    ...\n"
+            "  ]\n"
+            "}\n"
+        )
+        return messages + [
+            AIMessage(content=(f"Current question: {question}\nExisting answer to challenge: {existing_answer}\n")),
+            AIMessage(content=instructions),
+        ]
+
+
+# ---------------------------------------------------------------------------------
+# 2) Strict Typed Steps for Pipeline
+# ---------------------------------------------------------------------------------
+
+
+class StepName(StrEnum):
+    """Enum for step names in the pipeline."""
+
+    DOMAIN_DETECTION = "DomainDetection"
+    DECOMPOSITION = "Decomposition"
+    DECOMPOSITION_CRITIQUE = "DecompositionCritique"
+    CONTRACTED_QUESTION = "ContractedQuestion"
+    CONTRACTED_DIRECT_ANSWER = "ContractedDirectAnswer"
+    CONTRACT_CRITIQUE = "ContractCritique"
+    BEST_APPROACH_DECISION = "BestApproachDecision"
+    ENSEMBLE = "Ensemble"
+    FINAL_ANSWER = "FinalAnswer"
+
+    DEVILS_ADVOCATE = "DevilsAdvocate"
+    DEVILS_ADVOCATE_CRITIQUE = "DevilsAdvocateCritique"
+
+
+class StepRelation(StrEnum):
+    """Enum for relationship types in the reasoning graph."""
+
+    CRITIQUES = "CRITIQUES"
+    SELECTS = "SELECTS"
+    RESULT_OF = "RESULT_OF"
+    SPLIT_INTO = "SPLIT_INTO"
+    DEPEND_ON = "DEPEND_ON"
+    PRECEDES = "PRECEDES"
+    DECOMPOSED_BY = "DECOMPOSED_BY"
+
+
+class StepRecord(BaseModel):
+    """A typed record for each pipeline step."""
+
+    step_name: StepName
+    domain: Optional[str] = None
+    score: Optional[float] = None
+    used: Optional[StepName] = None
+    sub_questions: Optional[list[SubQuestionNode]] = None
+    parent_decomp_step_idx: Optional[int] = None
+    parent_subq_idx: Optional[int] = None
+    question: Optional[str] = None
+    thought: Optional[str] = None
+    answer: Optional[str] = None
+
+    def as_properties(self) -> dict[str, str | float | int | None]:
+        """Converts the StepRecord to a dictionary of properties."""
+        result: dict[str, str | float | int | None] = {}
+        if self.score is not None:
+            result["score"] = self.score
+        if self.domain:
+            result["domain"] = self.domain
+        if self.question:
+            result["question"] = self.question
+        if self.thought:
+            result["thought"] = self.thought
+        if self.answer:
+            result["answer"] = self.answer
+        return result
+
+
+# ---------------------------------------------------------------------------------
+# 3) Logging Setup
+# ---------------------------------------------------------------------------------
+
+
+class SimpleColorFormatter(logging.Formatter):
+    """Simple color-coded logging formatter for console output using ANSI escape codes."""
+
+    BLUE = "\033[94m"
+    GREEN = "\033[92m"
+    YELLOW = "\033[93m"
+    RED = "\033[91m"
+    RESET = "\033[0m"
+    LEVEL_COLORS = {
+        logging.DEBUG: BLUE,
+        logging.INFO: GREEN,
+        logging.WARNING: YELLOW,
+        logging.ERROR: RED,
+        logging.CRITICAL: RED,
+    }
+
+    def format(self, record: logging.LogRecord) -> str:
+        log_color = self.LEVEL_COLORS.get(record.levelno, self.RESET)
+        message = super().format(record)
+        return f"{log_color}{message}{self.RESET}"
+
+
+logger = logging.getLogger("AoT")
+logger.setLevel(logging.INFO)
+handler = logging.StreamHandler()
+handler.setFormatter(SimpleColorFormatter("%(levelname)s: %(message)s"))
+logger.handlers = [handler]
+logger.propagate = False
+
+
+# ---------------------------------------------------------------------------------
+# 4) The AoTPipeline Class (now with recursive devil's advocate at each sub-question)
+# ---------------------------------------------------------------------------------
+
+T = TypeVar(
+    "T",
+    bound=EnsembleResponse
+    | ContractQuestionResponse
+    | LabelResponse
+    | CritiqueResponse
+    | RecursiveDecomposeResponse
+    | DevilsAdvocateResponse,
+)
 
 
 @dataclass
 class AoTPipeline:
     """
-    Implements an Atom-of-Thought pipeline with:
-    1) Domain detection
-    2) Direct solution
-    3) Recursive Decomposition
-    4) Label step to refine sub-questions
-    5) Contract question
-    6) Contracted direct solution
-    7) Ensemble
-    8) (Optional) Score final answer if ground-truth is available
+    The pipeline orchestrates:
+      1) Recursive decomposition
+      2) For each sub-question, it tries a main approach + a devil's advocate approach
+      3) Merges sub-answers using an ensemble
+      4) Contracts the question
+      5) Possibly does a direct approach on the contracted question
+      6) Ensembling the final answers
     """
 
     chatterer: Chatterer
     max_depth: int = 2
     max_retries: int = 2
+    steps_history: list[StepRecord] = field(default_factory=list)
+    prompter: AoTPrompter = field(default_factory=AoTPrompter)
 
-    prompter_map: dict[Domain, BaseAoTPrompter] = field(
-        default_factory=lambda: {
-            Domain.GENERAL: GeneralAoTPrompter(),
-            Domain.MATH: MathAoTPrompter(),
-            Domain.CODING: CodingAoTPrompter(),
-            Domain.PHILOSOPHY: PhilosophyAoTPrompter(),
-            Domain.MULTIHOP: MultiHopAoTPrompter(),
-        }
-    )
-
-    async def _ainvoke_pydantic(self, prompt: str, model_cls: Type[T], default_answer: str = "Unable to process") -> T:
+    # 4.1) Utility for calling the LLM with Pydantic parsing
+    async def _ainvoke_pydantic(
+        self,
+        messages: list[BaseMessage],
+        model_cls: Type[T],
+        fallback: str = "<None>",
+    ) -> T:
         """
-        Attempts up to max_retries to parse the model_cls from LLM output.
+        Attempts up to max_retries to parse the model_cls from LLM output as JSON.
         """
         for attempt in range(1, self.max_retries + 1):
             try:
-                result = await self.chatterer.agenerate_pydantic(
-                    response_model=model_cls, messages=[{"role": "user", "content": prompt}]
-                )
-                logger.debug(f"[Validated attempt {attempt}] => {result.model_dump_json(indent=2)}")
-                return result
+                return await self.chatterer.agenerate_pydantic(response_model=model_cls, messages=messages)
             except ValidationError as e:
-                logger.warning(f"Validation error on attempt {attempt}/{self.max_retries}: {str(e)}")
+                logger.warning(f"ValidationError on attempt {attempt} for {model_cls.__name__}: {e}")
                 if attempt == self.max_retries:
-                    # Return an empty or fallback
-                    if model_cls == EnsembleResponse:
-                        return model_cls(thought="Failed label parse", answer=default_answer, confidence=0.0)
-                    elif model_cls == ContractQuestionResponse:
-                        return model_cls(thought="Failed contract parse", question="Unknown")
-                    elif model_cls == LabelResponse:
-                        return model_cls(thought="Failed label parse", sub_questions=[])
+                    # Return a fallback version
+                    if issubclass(model_cls, EnsembleResponse):
+                        return model_cls(thought=fallback, answer=fallback, confidence=0.0)  # type: ignore
+                    elif issubclass(model_cls, ContractQuestionResponse):
+                        return model_cls(thought=fallback, question=fallback)  # type: ignore
+                    elif issubclass(model_cls, LabelResponse):
+                        return model_cls(thought=fallback, sub_questions=[])  # type: ignore
+                    elif issubclass(model_cls, CritiqueResponse):
+                        return model_cls(thought=fallback, self_assessment=0.0)  # type: ignore
+                    elif issubclass(model_cls, DevilsAdvocateResponse):
+                        return model_cls(thought=fallback, final_answer=fallback, sub_questions=[])  # type: ignore
                     else:
-                        # fallback for Direct/Decompose
-                        return model_cls(thought="Failed parse", answer=default_answer)
+                        return model_cls(thought=fallback, final_answer=fallback, sub_questions=[])  # type: ignore
+        # theoretically unreachable
         raise RuntimeError("Unexpected error in _ainvoke_pydantic")
 
-    async def _adetect_domain(self, question: str, context: Optional[str] = None) -> Domain:
+    # 4.2) Helper method for self-critique
+    async def _ainvoke_critique(
+        self,
+        messages: list[BaseMessage],
+        thought: str,
+        answer: str,
+    ) -> CritiqueResponse:
         """
-        Queries an LLM to figure out which domain is best suited.
+        Instructs the LLM to critique the given thought & answer, returning CritiqueResponse.
         """
+        return await self._ainvoke_pydantic(
+            messages=self.prompter.critique_prompt(messages=messages, thought=thought, answer=answer),
+            model_cls=CritiqueResponse,
+        )
 
-        class InferredDomain(BaseModel):
-            domain: Domain
-
-        ctx_str = f"\nCONTEXT:\n{context}" if context else ""
-        domain_prompt = (
-            "You are an expert domain classifier. "
-            "Possible domains: [general, math, coding, philosophy, multihop].\n\n"
-            "Return valid JSON: {'domain': '...'}.\n"
-            f"QUESTION:\n{question}{ctx_str}"
+    # 4.3) Helper method for devil's advocate approach
+    async def _ainvoke_devils_advocate(
+        self,
+        messages: list[BaseMessage],
+        question: str,
+        existing_answer: str,
+    ) -> DevilsAdvocateResponse:
+        """
+        Instructs the LLM to challenge an existing answer with a devil's advocate approach.
+        """
+        return await self._ainvoke_pydantic(
+            messages=self.prompter.devils_advocate_prompt(messages, question=question, existing_answer=existing_answer),
+            model_cls=DevilsAdvocateResponse,
         )
-        try:
-            result: InferredDomain = await self.chatterer.agenerate_pydantic(
-                response_model=InferredDomain, messages=[{"role": "user", "content": domain_prompt}]
-            )
-            return result.domain
-        except ValidationError:
-            logger.warning("Failed domain detection, defaulting to general.")
-            return Domain.GENERAL
 
+    # 4.4) The main function that recursively decomposes a question and calls sub-steps
     async def _arecursive_decompose_question(
-        self, question: str, depth: int, prompter: BaseAoTPrompter, context: Optional[str] = None
+        self,
+        messages: list[BaseMessage],
+        question: str,
+        depth: int,
+        parent_decomp_step_idx: Optional[int] = None,
+        parent_subq_idx: Optional[int] = None,
     ) -> RecursiveDecomposeResponse:
         """
-        Recursively decomposes a question into sub-questions, applying an optional label step.
+        Recursively decompose the given question. For each sub-question:
+          1) Recursively decompose that sub-question if we still have depth left
+          2) After getting a main sub-answer, do a devil's advocate pass
+          3) Combine main sub-answer + devil's advocate alternative via an ensemble
         """
         if depth < 0:
-            return RecursiveDecomposeResponse(thought="Max depth reached", final_answer="Unknown", sub_questions=[])
+            logger.info("Max depth reached, returning unknown.")
+            return RecursiveDecomposeResponse(thought=MAX_DEPTH_REACHED, final_answer=UNKNOWN, sub_questions=[])
 
-        indent: LiteralString = "  " * (self.max_depth - depth)
-        logger.debug(f"{indent}Decomposing at depth {self.max_depth - depth}: {question}")
-
-        # Step 1: Base decomposition
-        prompt = prompter.recursive_decompose_prompt(question, context=context)
-        decompose_resp: RecursiveDecomposeResponse = await self._ainvoke_pydantic(prompt, RecursiveDecomposeResponse)
+        # Step 1: Perform the decomposition
+        decompose_resp: RecursiveDecomposeResponse = await self._ainvoke_pydantic(
+            messages=self.prompter.recursive_decompose_prompt(messages=messages, question=question, sub_answers=[]),
+            model_cls=RecursiveDecomposeResponse,
+        )
 
-        # Step 2: Label step to refine sub-questions if any
+        # Step 2: Label / refine sub-questions (dependencies, ordering)
         if decompose_resp.sub_questions:
-            label_prompt: str = prompter.label_prompt(question, decompose_resp, context=context)
-            label_resp: LabelResponse = await self._ainvoke_pydantic(label_prompt, LabelResponse)
-            # Overwrite the sub-questions with refined ones
+            label_resp: LabelResponse = await self._ainvoke_pydantic(
+                messages=self.prompter.label_prompt(messages, question, decompose_resp),
+                model_cls=LabelResponse,
+            )
             decompose_resp.sub_questions = label_resp.sub_questions
 
-        # Step 3: If depth > 0, try to resolve sub-questions recursively
-        #         so we can potentially update final_answer
+        # Save a pipeline record for this decomposition step
+        current_decomp_step_idx = self._record_decomposition_step(
+            question=question,
+            final_answer=decompose_resp.final_answer,
+            sub_questions=decompose_resp.sub_questions,
+            parent_decomp_step_idx=parent_decomp_step_idx,
+            parent_subq_idx=parent_subq_idx,
+        )
+
+        # Step 3: If sub-questions exist and depth remains, solve them + do devil's advocate
         if depth > 0 and decompose_resp.sub_questions:
-            resolved_subs: list[SubQuestionNode] = await self._aresolve_sub_questions(
-                decompose_resp.sub_questions, depth, prompter, context
+            solved_subs: list[SubQuestionNode] = await self._aresolve_sub_questions(
+                messages=messages,
+                sub_questions=decompose_resp.sub_questions,
+                depth=depth,
+                parent_decomp_step_idx=current_decomp_step_idx,
             )
-            # Step 4: Re-invoke decomposition with known sub-answers (sub_answers)
-            sub_answers_str: str = "\n".join(f"{sq.question}: {sq.answer}" for sq in resolved_subs if sq.answer)
-            if sub_answers_str:
-                refine_prompt: str = prompter.recursive_decompose_prompt(question, sub_answers_str, context=context)
-                refined_resp: RecursiveDecomposeResponse = await self._ainvoke_pydantic(
-                    refine_prompt, RecursiveDecomposeResponse
-                )
-                # Use the refined final answer, keep resolved sub-questions
-                decompose_resp.final_answer = refined_resp.final_answer
-                decompose_resp.sub_questions = resolved_subs
+            # Then we can refine the "final_answer" from those sub-answers
+            # or we do a secondary pass to refine the final answer
+            refined_prompt = self.prompter.recursive_decompose_prompt(
+                messages=messages,
+                question=question,
+                sub_answers=[(sq.question, sq.answer or UNKNOWN) for sq in solved_subs],
+            )
+            refined_resp: RecursiveDecomposeResponse = await self._ainvoke_pydantic(
+                refined_prompt, RecursiveDecomposeResponse
+            )
+            decompose_resp.final_answer = refined_resp.final_answer
+            decompose_resp.sub_questions = solved_subs
+
+            # Update pipeline record
+            self.steps_history[current_decomp_step_idx].answer = refined_resp.final_answer
+            self.steps_history[current_decomp_step_idx].sub_questions = solved_subs
 
         return decompose_resp
 
+    def _record_decomposition_step(
+        self,
+        question: str,
+        final_answer: str,
+        sub_questions: list[SubQuestionNode],
+        parent_decomp_step_idx: Optional[int],
+        parent_subq_idx: Optional[int],
+    ) -> int:
+        """
+        Save the decomposition step in steps_history, returning the index.
+        """
+        step_record = StepRecord(
+            step_name=StepName.DECOMPOSITION,
+            question=question,
+            answer=final_answer,
+            sub_questions=sub_questions,
+            parent_decomp_step_idx=parent_decomp_step_idx,
+            parent_subq_idx=parent_subq_idx,
+        )
+        self.steps_history.append(step_record)
+        return len(self.steps_history) - 1
+
     async def _aresolve_sub_questions(
-        self, sub_questions: list[SubQuestionNode], depth: int, prompter: BaseAoTPrompter, context: Optional[str] = None
+        self,
+        messages: list[BaseMessage],
+        sub_questions: list[SubQuestionNode],
+        depth: int,
+        parent_decomp_step_idx: Optional[int],
     ) -> list[SubQuestionNode]:
         """
-        Resolves each sub-question (potentially reusing the same decomposition approach)
-        in a topological order of dependencies.
+        Resolve sub-questions in topological order.
+        For each sub-question:
+          1) Recursively decompose (main approach).
+          2) Acquire a devil's advocate alternative.
+          3) Critique or ensemble if needed.
+          4) Finalize sub-question answer.
         """
-        n: int = len(sub_questions)
-        resolved: dict[int, SubQuestionNode] = {}
-
-        # Build adjacency
-        in_degree: list[int] = [0] * n
+        n = len(sub_questions)
+        in_degree = [0] * n
         graph: list[list[int]] = [[] for _ in range(n)]
         for i, sq in enumerate(sub_questions):
             for dep in sq.depend:
@@ -426,169 +560,417 @@ class AoTPipeline:
                     in_degree[i] += 1
                     graph[dep].append(i)
 
-        # Topological BFS
-        queue: list[int] = [i for i in range(n) if in_degree[i] == 0]
-        order: list[int] = []
+        # Kahn's algorithm for topological order
+        queue = [i for i in range(n) if in_degree[i] == 0]
+        topo_order: list[int] = []
+
         while queue:
             node = queue.pop(0)
-            order.append(node)
+            topo_order.append(node)
             for nxt in graph[node]:
                 in_degree[nxt] -= 1
                 if in_degree[nxt] == 0:
                     queue.append(nxt)
 
-        # If there's a cycle, some sub-questions won't appear in order
-        # but we'll attempt to resolve what we can
-        async def resolve_single_subq(idx: int) -> None:
-            sq: SubQuestionNode = sub_questions[idx]
-            # Attempt to answer this sub-question by decomposition if needed
-            sub_decomp: RecursiveDecomposeResponse = await self._arecursive_decompose_question(
-                sq.question, depth - 1, prompter, context
+        # We'll store the resolved sub-questions
+        final_subs: dict[int, SubQuestionNode] = {}
+
+        async def _resolve_one_subq(idx: int):
+            sq = sub_questions[idx]
+            # 1) Main approach
+            main_resp = await self._arecursive_decompose_question(
+                messages=messages,
+                question=sq.question,
+                depth=depth - 1,
+                parent_decomp_step_idx=parent_decomp_step_idx,
+                parent_subq_idx=idx,
             )
-            sq.answer = sub_decomp.final_answer
-            resolved[idx] = sq
 
-        await asyncio.gather(*(resolve_single_subq(i) for i in order))
+            main_answer = main_resp.final_answer
 
-        # Return only resolved sub-questions
-        return [resolved[i] for i in range(n) if i in resolved]
+            # 2) Devil's Advocate approach
+            devils_resp = await self._ainvoke_devils_advocate(
+                messages=messages, question=sq.question, existing_answer=main_answer
+            )
+            # 3) Ensemble to combine main_answer + devils_alternative
+            ensemble_sub = await self._ainvoke_pydantic(
+                self.prompter.ensemble_prompt(
+                    messages=messages,
+                    possible_thought_and_answers=[
+                        (main_resp.thought, main_answer),
+                        (devils_resp.thought, devils_resp.final_answer),
+                    ],
+                ),
+                EnsembleResponse,
+            )
+            sub_best_answer = ensemble_sub.answer
+
+            # Store final subq answer
+            sq.answer = sub_best_answer
+            final_subs[idx] = sq
+
+            # Record pipeline steps for devil's advocate
+            self.steps_history.append(
+                StepRecord(
+                    step_name=StepName.DEVILS_ADVOCATE,
+                    question=sq.question,
+                    answer=devils_resp.final_answer,
+                    thought=devils_resp.thought,
+                    sub_questions=devils_resp.sub_questions,
+                )
+            )
+            # Possibly critique the devils advocate result
+            dev_adv_crit = await self._ainvoke_critique(
+                messages=messages, thought=devils_resp.thought, answer=devils_resp.final_answer
+            )
+            self.steps_history.append(
+                StepRecord(
+                    step_name=StepName.DEVILS_ADVOCATE_CRITIQUE,
+                    thought=dev_adv_crit.thought,
+                    score=dev_adv_crit.self_assessment,
+                )
+            )
 
-    def _calculate_score(self, answer: str, ground_truth: Optional[str], domain: Domain) -> float:
-        """
-        Example scoring function. Real usage depends on having ground-truth.
-        If ground_truth is None, returns -1.0 as 'no score possible'.
+        # Solve sub-questions in topological order
+        tasks = [_resolve_one_subq(i) for i in topo_order]
+        await asyncio.gather(*tasks, return_exceptions=False)
 
-        This function is used for `post-hoc` scoring of the final answer.
-        """
-        if ground_truth is None:
-            return -1.0
+        return [final_subs[i] for i in range(n)]
 
-        # Very simplistic example:
-        # MATH: attempt numeric equality
-        if domain == Domain.MATH:
-            try:
-                ans_val = float(answer.strip())
-                gt_val = float(ground_truth.strip())
-                return 1.0 if abs(ans_val - gt_val) < 1e-9 else 0.0
-            except ValueError:
-                return 0.0
-
-        # For anything else, do a naive exact-match check ignoring case
-        return 1.0 if answer.strip().lower() == ground_truth.strip().lower() else 0.0
-
-    async def arun_pipeline(
-        self, question: str, context: Optional[str] = None, ground_truth: Optional[str] = None
-    ) -> str:
+    # 4.5) The primary pipeline method
+    async def arun_pipeline(self, messages: list[BaseMessage]) -> str:
         """
-        Full AoT pipeline. If ground_truth is provided, we compute a final score.
+        Execute the pipeline:
+          1) Decompose the main question (recursively).
+          2) Self-critique.
+          3) Provide a devil's advocate approach on the entire main result.
+          4) Contract sub-answers (optional).
+          5) Directly solve the contracted question.
+          6) Self-critique again.
+          7) Final ensemble across main vs devil's vs contracted direct answer.
+          8) Return final answer.
         """
-        # 1) Domain detection
-        domain = await self._adetect_domain(question, context)
-        prompter = self.prompter_map.get(domain, GeneralAoTPrompter())
-        logger.debug(f"Detected domain: {domain}")
-
-        # 2) Direct approach
-        direct_prompt = prompter.direct_prompt(question, context)
-        direct_resp = await self._ainvoke_pydantic(direct_prompt, DirectResponse)
-        direct_answer = direct_resp.answer
-        logger.debug(f"Direct answer => {direct_answer}")
-
-        # 3) Recursive Decomposition + label
-        decomp_resp = await self._arecursive_decompose_question(question, self.max_depth, prompter, context)
-        decompose_answer = decomp_resp.final_answer
-        logger.debug(f"Decomposition answer => {decompose_answer}")
-
-        # 4) Contract question
-        sub_answers_str = "\n".join(f"{sq.question}: {sq.answer}" for sq in decomp_resp.sub_questions if sq.answer)
-        contract_prompt = prompter.contract_prompt(question, sub_answers_str, context)
-        contract_resp = await self._ainvoke_pydantic(contract_prompt, ContractQuestionResponse)
+        self.steps_history.clear()
+
+        original_question: str = messages[-1].text()
+        # 1) Recursive decomposition
+        decomp_resp = await self._arecursive_decompose_question(
+            messages=messages,
+            question=original_question,
+            depth=self.max_depth,
+        )
+        logger.info(f"[Main Decomposition] final_answer={decomp_resp.final_answer}")
+
+        # 2) Self-critique of main decomposition
+        decomp_critique = await self._ainvoke_critique(
+            messages=messages, thought=decomp_resp.thought, answer=decomp_resp.final_answer
+        )
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.DECOMPOSITION_CRITIQUE,
+                thought=decomp_critique.thought,
+                score=decomp_critique.self_assessment,
+            )
+        )
+
+        # 3) Devil's advocate on the entire main answer
+        devils_on_main = await self._ainvoke_devils_advocate(
+            messages=messages, question=original_question, existing_answer=decomp_resp.final_answer
+        )
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.DEVILS_ADVOCATE,
+                question=original_question,
+                answer=devils_on_main.final_answer,
+                thought=devils_on_main.thought,
+                sub_questions=devils_on_main.sub_questions,
+            )
+        )
+        devils_crit_main = await self._ainvoke_critique(
+            messages=messages, thought=devils_on_main.thought, answer=devils_on_main.final_answer
+        )
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.DEVILS_ADVOCATE_CRITIQUE,
+                thought=devils_crit_main.thought,
+                score=devils_crit_main.self_assessment,
+            )
+        )
+
+        # 4) Contract sub-answers from main decomposition
+        top_decomp_record: Optional[StepRecord] = next(
+            (
+                s
+                for s in reversed(self.steps_history)
+                if s.step_name == StepName.DECOMPOSITION and s.parent_decomp_step_idx is None
+            ),
+            None,
+        )
+        if top_decomp_record and top_decomp_record.sub_questions:
+            sub_answers = [(sq.question, sq.answer or UNKNOWN) for sq in top_decomp_record.sub_questions]
+        else:
+            sub_answers = []
+
+        contract_resp = await self._ainvoke_pydantic(
+            messages=self.prompter.contract_prompt(messages, sub_answers),
+            model_cls=ContractQuestionResponse,
+        )
         contracted_question = contract_resp.question
-        logger.debug(f"Contracted question => {contracted_question}")
-
-        # 5) Direct approach on contracted question
-        contracted_direct_prompt = prompter.direct_prompt(contracted_question, context)
-        contracted_direct_resp = await self._ainvoke_pydantic(contracted_direct_prompt, DirectResponse)
-        contracted_direct_answer = contracted_direct_resp.answer
-        logger.debug(f"Contracted direct answer => {contracted_direct_answer}")
-
-        # 6) Ensemble
-        ensemble_prompt = prompter.ensemble_prompt(
-            original_question=question,
-            direct_answer=direct_answer,
-            decompose_answer=decompose_answer,
-            contracted_direct_answer=contracted_direct_answer,
-            context=context,
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.CONTRACTED_QUESTION, question=contracted_question, thought=contract_resp.thought
+            )
         )
-        ensemble_resp = await self._ainvoke_pydantic(ensemble_prompt, EnsembleResponse)
-        final_answer = ensemble_resp.answer
-        logger.debug(f"Ensemble final answer => {final_answer} (confidence={ensemble_resp.confidence})")
 
-        # 7) Optional scoring
-        final_score = self._calculate_score(final_answer, ground_truth, domain)
-        if final_score >= 0.0:
-            logger.info(f"Final Score: {final_score:.3f} (domain={domain})")
+        # 5) Attempt direct approach on contracted question
+        contracted_direct = await self._ainvoke_pydantic(
+            messages=self.prompter.contract_direct_prompt(messages, contracted_question),
+            model_cls=RecursiveDecomposeResponse,
+            fallback="No Contracted Direct Answer",
+        )
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.CONTRACTED_DIRECT_ANSWER,
+                answer=contracted_direct.final_answer,
+                thought=contracted_direct.thought,
+            )
+        )
+        logger.info(f"[Contracted Direct] final_answer={contracted_direct.final_answer}")
 
-        return final_answer
+        # 5.1) Critique the contracted direct approach
+        contract_critique = await self._ainvoke_critique(
+            messages=messages, thought=contracted_direct.thought, answer=contracted_direct.final_answer
+        )
+        self.steps_history.append(
+            StepRecord(
+                step_name=StepName.CONTRACT_CRITIQUE,
+                thought=contract_critique.thought,
+                score=contract_critique.self_assessment,
+            )
+        )
 
+        # 6) Ensemble of (Main decomposition, Devil's advocate on main, Contracted direct)
+        ensemble_resp = await self._ainvoke_pydantic(
+            self.prompter.ensemble_prompt(
+                messages=messages,
+                possible_thought_and_answers=[
+                    (decomp_resp.thought, decomp_resp.final_answer),
+                    (devils_on_main.thought, devils_on_main.final_answer),
+                    (contracted_direct.thought, contracted_direct.final_answer),
+                ],
+            ),
+            EnsembleResponse,
+        )
+        best_approach_answer = ensemble_resp.answer
+        approach_used = StepName.ENSEMBLE
+        self.steps_history.append(StepRecord(step_name=StepName.BEST_APPROACH_DECISION, used=approach_used))
+        logger.info(f"[Best Approach Decision] => {approach_used}")
+
+        # 7) Final answer
+        self.steps_history.append(
+            StepRecord(step_name=StepName.FINAL_ANSWER, answer=best_approach_answer, score=ensemble_resp.confidence)
+        )
+        logger.info(f"[Final Answer] => {best_approach_answer}")
 
-# ------------------ 3) AoTStrategy that uses the pipeline ------------------
+        return best_approach_answer
 
+    def run_pipeline(self, messages: list[BaseMessage]) -> str:
+        """Synchronous wrapper around arun_pipeline."""
+        return asyncio.run(self.arun_pipeline(messages))
 
-@dataclass(kw_only=True)
-class AoTStrategy(BaseStrategy):
-    pipeline: AoTPipeline
+    # ---------------------------------------------------------------------------------
+    # 4.6) Build or export a reasoning graph
+    # ---------------------------------------------------------------------------------
 
-    async def ainvoke(self, messages: LanguageModelInput) -> str:
-        logger.debug(f"Invoking with messages: {messages}")
+    def get_reasoning_graph(self, global_id_prefix: str = "AoT"):
+        """
+        Constructs a Graph object (from hypothetical `neo4j_extension`)
+        capturing the pipeline steps, including devil's advocate steps.
+        """
+        from neo4j_extension import Graph, Node, Relationship
+
+        g = Graph()
+        step_nodes: dict[int, Node] = {}
+        subq_nodes: dict[str, Node] = {}
+
+        # Step A: Create nodes for each pipeline step
+        for i, record in enumerate(self.steps_history):
+            # We'll skip nested Decomposition steps only if we want to flatten them.
+            # But let's keep them for clarity.
+            step_node = Node(
+                properties=record.as_properties(), labels={record.step_name}, globalId=f"{global_id_prefix}_step_{i}"
+            )
+            g.add_node(step_node)
+            step_nodes[i] = step_node
+
+        # Step B: Collect sub-questions from each DECOMPOSITION or DEVILS_ADVOCATE
+        all_sub_questions: dict[str, tuple[int, int, SubQuestionNode]] = {}
+        for i, record in enumerate(self.steps_history):
+            if record.sub_questions:
+                for sq_idx, sq in enumerate(record.sub_questions):
+                    sq_id = f"{global_id_prefix}_decomp_{i}_sub_{sq_idx}"
+                    all_sub_questions[sq_id] = (i, sq_idx, sq)
+
+        for sq_id, (i, sq_idx, sq) in all_sub_questions.items():
+            n_subq = Node(
+                properties={
+                    "question": sq.question,
+                    "answer": sq.answer or "",
+                },
+                labels={"SubQuestion"},
+                globalId=sq_id,
+            )
+            g.add_node(n_subq)
+            subq_nodes[sq_id] = n_subq
+
+        # Step C: Add relationships. We do a simple approach:
+        #  - If StepRecord is DECOMPOSITION or DEVILS_ADVOCATE with sub_questions, link them via SPLIT_INTO.
+        for i, record in enumerate(self.steps_history):
+            if record.sub_questions:
+                start_node = step_nodes[i]
+                for sq_idx, sq in enumerate(record.sub_questions):
+                    sq_id = f"{global_id_prefix}_decomp_{i}_sub_{sq_idx}"
+                    end_node = subq_nodes[sq_id]
+                    rel = Relationship(
+                        properties={},
+                        rel_type=StepRelation.SPLIT_INTO,
+                        start_node=start_node,
+                        end_node=end_node,
+                        globalId=f"{global_id_prefix}_split_{i}_{sq_idx}",
+                    )
+                    g.add_relationship(rel)
+                    # Also add sub-question dependencies
+                    for dep in sq.depend:
+                        # The same record i -> sub-question subq
+                        if 0 <= dep < len(record.sub_questions):
+                            dep_id = f"{global_id_prefix}_decomp_{i}_sub_{dep}"
+                            if dep_id in subq_nodes:
+                                dep_node = subq_nodes[dep_id]
+                                rel_dep = Relationship(
+                                    properties={},
+                                    rel_type=StepRelation.DEPEND_ON,
+                                    start_node=end_node,
+                                    end_node=dep_node,
+                                    globalId=f"{global_id_prefix}_dep_{i}_q_{sq_idx}_on_{dep}",
+                                )
+                                g.add_relationship(rel_dep)
+
+        # Step D: We add PRECEDES relationships in a linear chain for the pipeline steps
+        for i in range(len(self.steps_history) - 1):
+            start_node = step_nodes[i]
+            end_node = step_nodes[i + 1]
+            rel = Relationship(
+                properties={},
+                rel_type=StepRelation.PRECEDES,
+                start_node=start_node,
+                end_node=end_node,
+                globalId=f"{global_id_prefix}_precede_{i}_to_{i + 1}",
+            )
+            g.add_relationship(rel)
+
+        # Step E: CRITIQUES, SELECTS, RESULT_OF can be similarly added:
+        # We'll do a simple pass:
+        # If step_name ends with CRITIQUE => it critiques the step before it
+        for i, record in enumerate(self.steps_history):
+            if "CRITIQUE" in record.step_name:
+                # Let it point to the preceding step
+                if i > 0:
+                    start_node = step_nodes[i]
+                    end_node = step_nodes[i - 1]
+                    rel = Relationship(
+                        properties={},
+                        rel_type=StepRelation.CRITIQUES,
+                        start_node=start_node,
+                        end_node=end_node,
+                        globalId=f"{global_id_prefix}_crit_{i}",
+                    )
+                    g.add_relationship(rel)
+
+        # If there's a BEST_APPROACH_DECISION step, link it to the step it uses
+        best_decision_idx = None
+        used_step_idx = None
+        for i, record in enumerate(self.steps_history):
+            if record.step_name == StepName.BEST_APPROACH_DECISION and record.used:
+                best_decision_idx = i
+                # find the step with that name
+                used_step_idx = next((j for j in step_nodes if self.steps_history[j].step_name == record.used), None)
+                if used_step_idx is not None:
+                    rel = Relationship(
+                        properties={},
+                        rel_type=StepRelation.SELECTS,
+                        start_node=step_nodes[i],
+                        end_node=step_nodes[used_step_idx],
+                        globalId=f"{global_id_prefix}_select_{i}_use_{used_step_idx}",
+                    )
+                    g.add_relationship(rel)
+
+        # And link the final answer to the best approach
+        final_answer_idx = next(
+            (i for i, r in enumerate(self.steps_history) if r.step_name == StepName.FINAL_ANSWER), None
+        )
+        if final_answer_idx is not None and best_decision_idx is not None:
+            rel = Relationship(
+                properties={},
+                rel_type=StepRelation.RESULT_OF,
+                start_node=step_nodes[final_answer_idx],
+                end_node=step_nodes[best_decision_idx],
+                globalId=f"{global_id_prefix}_final_{final_answer_idx}_resultof_{best_decision_idx}",
+            )
+            g.add_relationship(rel)
 
-        input = self.pipeline.chatterer.client._convert_input(messages)
-        input_string = input.to_string()
-        logger.debug(f"Extracted question: {input_string}")
-        return await self.pipeline.arun_pipeline(input_string)
+        return g
 
-    def invoke(self, messages: LanguageModelInput) -> str:
-        return asyncio.run(self.ainvoke(messages))
 
+# ---------------------------------------------------------------------------------
+# 5) AoTStrategy class that uses the pipeline
+# ---------------------------------------------------------------------------------
 
-# ------------------ 4) Example usage (main) ------------------
-if __name__ == "__main__":
-    from warnings import filterwarnings
 
-    import colorama
+@dataclass
+class AoTStrategy(BaseStrategy):
+    """
+    Strategy using AoTPipeline with a reasoning graph and deeper devil's advocate.
+    """
 
-    filterwarnings("ignore", category=UserWarning)
+    pipeline: AoTPipeline
 
-    colorama.init(autoreset=True)
+    async def ainvoke(self, messages: LanguageModelInput) -> str:
+        """Asynchronously run the pipeline with the given messages."""
+        # Convert your custom input to list[BaseMessage] as needed:
+        msgs = self.pipeline.chatterer.client._convert_input(messages).to_messages()  # type: ignore
+        return await self.pipeline.arun_pipeline(msgs)
 
-    class ColoredFormatter(logging.Formatter):
-        COLORS = {
-            "DEBUG": colorama.Fore.CYAN,
-            "INFO": colorama.Fore.GREEN,
-            "WARNING": colorama.Fore.YELLOW,
-            "ERROR": colorama.Fore.RED,
-            "CRITICAL": colorama.Fore.RED + colorama.Style.BRIGHT,
-        }
+    def invoke(self, messages: LanguageModelInput) -> str:
+        """Synchronously run the pipeline with the given messages."""
+        msgs = self.pipeline.chatterer.client._convert_input(messages).to_messages()  # type: ignore
+        return self.pipeline.run_pipeline(msgs)
 
-        def format(self, record):
-            levelname = record.levelname
-            message = super().format(record)
-            return f"{self.COLORS.get(levelname, colorama.Fore.WHITE)}{message}{colorama.Style.RESET_ALL}"
+    def get_reasoning_graph(self):
+        """Return the AoT reasoning graph from the pipeline’s steps history."""
+        return self.pipeline.get_reasoning_graph(global_id_prefix="AoT")
 
-    logger.setLevel(logging.DEBUG)
-    handler = logging.StreamHandler()
-    handler.setLevel(logging.DEBUG)
-    formatter = ColoredFormatter("%(asctime)s - %(name)s - %(levelname)s - %(message)s")
-    handler.setFormatter(formatter)
-    logger.addHandler(handler)
 
-    file_handler = logging.FileHandler("atom_of_thoughts.log", encoding="utf-8", mode="w")
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(file_handler)
-    logger.propagate = False
+# ---------------------------------------------------------------------------------
+# Example usage (pseudo-code)
+# ---------------------------------------------------------------------------------
+if __name__ == "__main__":
+    from neo4j_extension import Neo4jConnection  # or your actual DB connector
 
-    pipeline = AoTPipeline(chatterer=Chatterer.openai(), max_depth=2)
+    # You would create a Chatterer with your chosen LLM backend (OpenAI, etc.)
+    chatterer = Chatterer.openai()  # pseudo-code
+    pipeline = AoTPipeline(chatterer=chatterer, max_depth=3)
     strategy = AoTStrategy(pipeline=pipeline)
 
-    question = "What would Newton discover if hit by an apple falling from 100 meters?"
+    question = "Solve 5.9 = 5.11 - x. Also compare 9.11 and 9.9."
     answer = strategy.invoke(question)
-    print(answer)
+    print("Final Answer:", answer)
+
+    # Build the reasoning graph
+    graph = strategy.get_reasoning_graph()
+    print(f"\nGraph has {len(graph.nodes)} nodes and {len(graph.relationships)} relationships.")
+
+    # Optionally store in Neo4j
+    with Neo4jConnection() as conn:
+        conn.clear_all()
+        conn.upsert_graph(graph)
+        print("Graph stored in Neo4j.")