app.py

import gradio as gr
import os
import asyncio
from llama_index.core.agent.workflow import AgentWorkflow
from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI

# ------------------
# 1. Tools
# ------------------
def add(a: int, b: int) -> int:
    """Adds two integers together."""
    return a + b

def subtract(a: int, b: int) -> int:
    """Subtracts one integer from another."""
    return a - b

# ------------------
# 2. Hugging Face LLM
# ------------------
llm = HuggingFaceInferenceAPI(
    model_name="zai-org/GLM-4.7",
    token=os.getenv("HF_TOKEN"),
    context_window=4096,
    max_new_tokens=512,
    generate_kwargs={
        "temperature": 0.1,
    },
)

# ------------------
# 3. Agent Setup (The Fixed Part)
# ------------------
# We use from_tools_or_functions to automatically create the agent logic
agent = AgentWorkflow.from_tools_or_functions(
    [add, subtract],
    llm=llm,
    system_prompt="You are a calculator assistant. Use tools for math calculations."
)

# ------------------
# 4. Streaming handler
# ------------------
async def chat_stream(user_msg):
    partial_text = ""
    # astream_run is the correct method for AgentWorkflow
    handler = agent.astream_run(user_msg=user_msg)
    
    async for event in handler:
        # This safely checks for text updates without needing specific imports
        if hasattr(event, "delta") and event.delta is not None:
            partial_text += event.delta
            yield partial_text

# ------------------
# 5. Gradio UI
# ------------------
demo = gr.Interface(
    fn=chat_stream,
    inputs=gr.Textbox(label="Ask", placeholder="e.g. What is 50 minus 20?"),
    outputs=gr.Textbox(label="Response"),
    title="HF LLM Calculator Agent",
)

demo.launch()