Lab 3: Chat Assistant & Customer Support Agent

Build two portfolio projects: a Python chat assistant with the Foundry SDK that remembers the conversation and follows a banking FAQ persona, and a customer support agent in the Foundry portal grounded in a policy document, called from a Python script that logs every exchange.

Time: 2–3 hours · Projects: #1 Foundry Chat Assistant, #2 Customer Support Agent · You need: the Lab 1 project and model deployment, Python 3.10+, the Azure CLI, Git.

Outcome. A working chat app and an agent, both pushed to your GitHub portfolio.

Step 1: Project setup

mkdir foundry-bank-assistant && cd foundry-bank-assistant
python -m venv .venv
# Windows: .venv\Scripts\activate    macOS/Linux: source .venv/bin/activate
pip install "azure-ai-projects>=2.3.0" azure-identity openai python-dotenv
az login
echo ".env"   >  .gitignore
echo ".venv/" >> .gitignore

Create .env:

FOUNDRY_PROJECT_ENDPOINT=https://<resource>.services.ai.azure.com/api/projects/<project>
FOUNDRY_MODEL_NAME=chat-model

If DefaultAzureCredential fails with a permissions error, open your project in the Azure portal → Access control (IAM) → Add role assignment → Azure AI User → your account.

Step 2: A chat client that remembers history

chat.py:

import os, time
from dotenv import load_dotenv
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from openai import RateLimitError

load_dotenv()
MODEL = os.environ["FOUNDRY_MODEL_NAME"]
project = AIProjectClient(endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"], credential=DefaultAzureCredential())
client = project.get_openai_client()

SYSTEM = """You are "Asha", the FAQ assistant for Contoso Bank India.
- Answer only questions about Contoso savings accounts, debit/credit cards and UPI.
- Keep answers under 80 words.
- Never ask for or repeat OTPs, PINs, CVV or full card numbers; warn the user if they share one.
- If you are unsure or the question is out of scope, say so and suggest calling 1800-000-000."""

history = [{"role": "system", "content": SYSTEM}]
MAX_TURNS = 10                                    # keep the context window small

def ask(text):
    history.append({"role": "user", "content": text})
    trimmed = [history[0]] + history[1:][-2 * MAX_TURNS:]
    for attempt in range(4):
        try:
            r = client.responses.create(model=MODEL, input=trimmed, max_output_tokens=300)
            history.append({"role": "assistant", "content": r.output_text})
            print(f"   (tokens: {r.usage.total_tokens})")
            return r.output_text
        except RateLimitError:
            time.sleep(2 ** attempt)
    return "Sorry, the service is busy. Please try again."

print("Asha is ready. Type 'quit' to exit.")
while (q := input("You: ").strip()).lower() not in {"quit", "exit"}:
    print("Asha:", ask(q))

Run python chat.py.

Step 3: Test the persona and edge cases

Write the results in TESTS.md. If a test fails, tighten the system prompt and rerun. This is how real teams iterate.

Step 4: Create the customer support agent in the portal

1. Create contoso-bank-policy.md (or a PDF) with a few sections, for example:

# Contoso Bank Card Policy (sample)
## Annual fees
Classic card: free. Gold card: Rs 499 (waived on spends above Rs 1 lakh per year). Platinum: Rs 1,499.
## Lost or stolen cards
Block immediately in the app or call 1800-000-000. Replacement card fee: Rs 200, delivered in 7 working days.
## Disputed transactions
Raise within 30 days of the statement date. Temporary credit within 10 working days after investigation starts.

2. In the Foundry portal: Agents → Create agent, name it support-agent, and choose your model deployment.
3. Paste the instructions:

You are Contoso Bank's customer support agent.
Use ONLY the attached card policy to answer. Mention the policy section you used.
If the policy does not answer the question, say "I'll raise this with our team" and summarise the issue in one line.
Never request OTPs, PINs, CVV or full card numbers.

4. Add knowledge → File search and upload the policy file.
5. Test in the playground: “What is the Gold card fee and can it be waived?”, “I lost my card in Chennai, what now?”, “Do you offer home loans?” Check the citations and trace.

Step 5: Call the agent from Python and log the conversation

agent_client.py:

import os, json, datetime
from dotenv import load_dotenv
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient

load_dotenv()
project = AIProjectClient(endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"], credential=DefaultAzureCredential())
openai = project.get_openai_client(agent_name="support-agent")
conversation = openai.conversations.create()

log_file = f"conversation-{datetime.date.today()}.jsonl"
questions = [
    "What is the annual fee for a Gold card?",
    "Can it be waived?",
    "I want to dispute a charge from 45 days ago.",
]
with open(log_file, "a", encoding="utf-8") as log:
    for q in questions:
        r = openai.responses.create(conversation=conversation.id, input=q)
        print(f"You:   {q}\nAgent: {r.output_text}\n")
        log.write(json.dumps({"time": datetime.datetime.now().isoformat(), "conversation": conversation.id,
                              "user": q, "agent": r.output_text}) + "\n")
print("Saved log to", log_file)

Step 6: Push to GitHub

git init && git add . && git commit -m "Foundry chat assistant and support agent (AI-901 Lab 3)"
git branch -M main
git remote add origin <your-new-repo-url>
git push -u origin main

Add a README.md with a description, an architecture line (Python → Foundry SDK → model deployment / agent with file search), setup steps, the test table and a screenshot. Check that .env is not in the repository.

Resume line: Built a banking FAQ chat assistant and a policy-grounded support agent on Microsoft Foundry (Python, Foundry SDK, file search), with conversation memory, rate-limit handling and logging.

Check yourself