building a self-correcting agent flow that can generate and iterate on 3D designs

[isayahc]

very simple langchain agent that is able to build 3d objects (warning not, currently multimodal and is not self correcting and uses cadquery to generate the step file )

from langchain_community.chat_models import ChatLiteLLM
from langchain.agents import create_agent
import os
import cadquery as cq
from langchain_core.tools import tool
import sys

@tool
def generate_step_box(length: float, width: float, height: float, filename: str = "agent_part.step") -> str:
    """
    Generates a 3D CAD box with the specified dimensions and exports it as a STEP file.
    
    Args:
        length (float): The length of the box along the X axis.
        width (float): The width of the box along the Y axis.
        height (float): The height of the box along the Z axis.
        filename (str): The name of the output STEP file. Defaults to 'agent_part.step'.
        
    Returns:
        str: A message indicating success or failure of the file generation.
    """
    try:
        # Generate the geometry using CadQuery
        model = cq.Workplane("XY").box(length, width, height)
        
        # Export the model to a working STEP file
        cq.exporters.export(model, filename)
        
        return f"Successfully generated CAD model and saved to {filename} with dimensions L:{length}, W:{width}, H:{height}."
    except Exception as e:
        return f"Failed to generate CAD model. Error: {str(e)}"

# 1. Put the tool in a list
tools = [generate_step_box]


def build_agent():
    """Build and return the agent graph."""
    # llm = ChatOpenAI(model=MODEL, temperature=TEMPERATURE)
    llm = ChatLiteLLM(model="anthropic/claude-3-haiku-20240307", temperature=0.1)

    return create_agent(llm,tools=tools, system_prompt="build stuff please")
    # return create_react_agent(llm, tools=TOOLS, prompt=SYSTEM_PROMPT)

def run_agent(question: str) -> str:
    """Run the agent on a single question and return the response text."""
    return run_agent_with_tools(question)["response"]


def run_agent_with_tools(question: str) -> dict:
    """Run the agent and return response + tool usage info for evaluation."""
    agent = build_agent()
    result = agent.invoke({"messages": [{"role": "user", "content": question}]})
    messages = result["messages"]

    tools_used = []
    for msg in messages:
        if hasattr(msg, "tool_calls") and msg.tool_calls:
            for tc in msg.tool_calls:
                tools_used.append(tc["name"])

    return {
        "response": messages[-1].content,
        "tools_used": tools_used,
    }

# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------

if __name__ == "__main__":
    if len(sys.argv) < 2:
        print("Usage: python agent.py <question>")
        sys.exit(1)
    question = " ".join(sys.argv[1:])
    print(run_agent(question))

sample output

References

• https://github.com/hwchase17/autoresearch-agents/blob/master/agent.py

[deanhu0822]

I don't want to include cq or cad query into OpenCAD, can you make it for just for ours