From LLMs to AI Agents — Why a Prompt Is Not Enough

import os
import asyncio
from pathlib import Path

from semantic_kernel.agents import ChatCompletionAgent
from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion, OpenAIChatPromptExecutionSettings
from semantic_kernel.functions import kernel_function, KernelArguments

# --- Configuration ---
from dotenv import load_dotenv
load_dotenv()

AZURE_OPENAI_ENDPOINT = os.environ["AZURE_OPENAI_ENDPOINT"]
AZURE_OPENAI_KEY = os.environ["AZURE_OPENAI_API_KEY"]
AZURE_OPENAI_DEPLOYMENT = os.environ["AZURE_OPENAI_DEPLOYMENT"]  # e.g. "gpt-4o-mini"

FAQ_DIR = Path("../data/faq_docs")

def load_faq_docs() -> dict[str, str]:
    """Load each .txt file into a {key: content} dictionary."""
    docs = {}
    for path in FAQ_DIR.glob("*.txt"):
        docs[path.stem] = path.read_text(encoding="utf-8").strip()
    return docs

FAQ = load_faq_docs()  # Loaded once at startup

# --- Tool registration ---
class InternalFaqTool:
    @kernel_function(
        name="lookup_faq",
        description="Lookup an internal policy document by key. Valid keys: 'release-freeze', 'incident-sev1'.",
    )
    def lookup_faq(self, key: str) -> str:
        return FAQ.get(key, "Policy not found. Please check with your Release Manager.")

# --- System prompt ---
instructions = """
You are a helpful internal knowledge assistant.
If you need a policy, call the lookup_faq tool.

You can answer questions about:
- Release freeze timelines and what is allowed during the freeze
- SEV1 incident handling, roles, escalation, and communication
"""

# --- LLM parameters ---
SETTINGS = OpenAIChatPromptExecutionSettings(
    temperature=0.2,
    max_tokens=500,
    tool_choice="auto",
)

# --- Model configuration ---
_agent = ChatCompletionAgent(
    service=AzureChatCompletion(
        deployment_name=AZURE_OPENAI_DEPLOYMENT,
        endpoint=AZURE_OPENAI_ENDPOINT,
        api_key=AZURE_OPENAI_KEY,
    ),
    name="Policy-Assistant",
    instructions=instructions,
    plugins=[InternalFaqTool()],
    arguments=KernelArguments(SETTINGS),
)

# --- Execution call ---
async def ask_agent(question: str) -> str:
    response = await _agent.get_response(messages=question)
    return response.content

async def main() -> None:
    user_question = input(
        "Ask a policy question: "
    ).strip()
    answer = await ask_agent(user_question)
    print(answer)

if __name__ == "__main__":
    asyncio.run(main())