| import json | |
| from typing import List, Optional | |
| from pydantic import BaseModel, Field | |
| from src.llm_factory import get_llm | |
| class PlanStep(BaseModel): | |
| """Details about a single step in the plan.""" | |
| step_title: str = Field( | |
| description="What does this step do briefly. Without formatted text." | |
| ) | |
| detailed_instructions: str = Field( | |
| description="What does this step do in great detail" | |
| ) | |
| class DocumentDetails(BaseModel): | |
| """A description of what problem needs to be solved.""" | |
| step_items: list[PlanStep] = Field( | |
| description="A list with the steps of the plan" | |
| ) | |
| llm = get_llm("ollama-llama3.1") | |
| sllm = llm.as_structured_llm(DocumentDetails) | |
| text = """ | |
| You are an expert at breaking down complex problems into smaller, more manageable parts. | |
| Create a rough plan for how to solve an ARC (Abstraction & Reasoning Corpus) puzzle. | |
| You have the input and output of the puzzle as a list of 2D arrays. | |
| Ideas to include in the plan: | |
| - Input/Output Analysis, look for clues within the input or output, are the shapes always square? | |
| - Hypothesis Generation, is this a sorting problem? are there things that has to be counted? | |
| - Refinement and Iteration | |
| - Document your thought process | |
| - Failure Analysis, did one of the step miss something crucial? | |
| - Solution Verification | |
| """ | |
| response = sllm.complete(text) | |
| json_response = json.loads(response.text) | |
| print(json.dumps(json_response, indent=2)) | |