embr-foundry-chat-sample-python

A tiny chat app that demonstrates Scenario 1 of the Embr × Foundry POC:

An agent is orchestrated inside an Embr-hosted Python app using Microsoft Agent Framework. The Foundry project only provides the underlying LLM (the model deployment) — the agent loop, system prompt, tool-calling, and memory all live in this repo's own code.

┌──────────────────────────────────────────────┐
│              Embr-hosted container           │
│                                              │
│   ┌─────────────┐     ┌──────────────────┐   │
│   │   FastAPI   │────▶│  Microsoft       │   │
│   │   + HTML UI │     │  Agent Framework │   │
│   └─────────────┘     │  (agent loop +   │   │
│                       │   local tools)   │   │
│                       └────────┬─────────┘   │
└────────────────────────────────┼─────────────┘
                                 │ inference only
                                 ▼
                      ┌──────────────────────┐
                      │  Foundry project     │
                      │  → model deployment  │
                      │  (gpt-4o / ...)      │
                      └──────────────────────┘

Two toy tools are wired in (get_weather, roll_dice) so you can actually see tool-calling happen. Replace them with whatever your demo needs.

---

Part 1 — Create the Foundry resources

You do this once, in the Azure AI Foundry portal. The app just needs a model endpoint + key.

1. Create a Foundry project

1. Go to ai.azure.com and sign in.
2. Click + New project.
3. Pick an existing hub or let the portal create one. Region: anywhere that has the model you want (e.g., East US 2).
4. After the project is provisioned, open it.

2. Deploy a model

1. In the left nav choose Models + endpoints.
2. Click + Deploy model → Deploy base model.
3. Pick gpt-4o-mini (cheap + fast for this demo) or gpt-4o.
4. Accept the default deployment name or give it one you'll remember (e.g., gpt-4o-mini).
5. Click Deploy.

3. Grab the three values you need

From your model deployment page in the Foundry portal, the Target URI will look like:

https://<resource>.services.ai.azure.com/api/projects/<project>/openai/v1/responses

You want the base of that URL (everything up to /openai/v1, dropping the /responses suffix). The OpenAI SDK will append the right path (/chat/completions) itself.

Env var	Where to find it	Example
FOUNDRY_BASE_URL	Target URI with the trailing /responses removed	https://jordan-embr.services.ai.azure.com/api/projects/proj-default/openai/v1
FOUNDRY_API_KEY	Keys + Endpoint on the Foundry resource (click the key icon to reveal)	abc123…
FOUNDRY_MODEL_DEPLOYMENT	The deployment name you chose when you deployed the model	gpt-5.4-mini-1

Resource-scoped endpoints (https://<resource>.openai.azure.com/openai/v1) also work if you prefer. Both point at the same underlying inference, just via different auth/routing paths.

---

Part 2 — Run locally

python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# fill in the three values from Part 1

uvicorn app.main:app --reload --port 8000

Open http://localhost:8000. Try:

• "What's the weather in Seattle and Tokyo?" → should call get_weather twice and summarize.
• "Roll 3d20 and tell me the highest." → should call roll_dice three times.

---

Part 3 — Deploy to Embr

Prerequisite

The Embr GitHub App must be installed on this repo.

Deploy

# Push this repo to GitHub first (under your account)
gh repo create embr-foundry-chat-sample-python --source=. --public --push

# One-command deploy
embr quickstart deploy <your-user>/embr-foundry-chat-sample-python -i <installation-id>

embr installations config returns the installation ID if you don't know it.

Configure the Foundry values

embr variables set FOUNDRY_BASE_URL https://<resource>.openai.azure.com/openai/v1
embr variables set FOUNDRY_API_KEY <your-key> --secret
embr variables set FOUNDRY_MODEL_DEPLOYMENT <deployment-name>

Trigger a redeploy so the new env vars take effect:

embr deployments create --restart

---

Project layout

.
├── app/
│   ├── __init__.py
│   ├── agent.py          # Agent definition + local tools (get_weather, roll_dice)
│   ├── main.py           # FastAPI: /, /health, /api/chat, /api/config
│   └── static/index.html # Minimal chat UI
├── embr.yaml             # platform: python 3.12, port 8000
├── requirements.txt
├── .env.example
└── README.md

Known limitations / gaps (feeding into the POC findings doc)

• No auth on /api/chat. V1 intentionally skipped auth so we could focus on the Embr ↔ Foundry wiring. Before shipping anything like this for real, put Entra in front of it.
• Conversation history is in-memory. Pod restart = lost context. Would want Redis / Cosmos in a real app.
• API-key auth to Foundry. Managed identity would be the secure story; not attempted in this POC.
• No streaming. agent.run() waits for the full response. Streaming via agent.run_stream() + SSE would be a small follow-up.