2-notebooks/1-chat_completion/4-phi-4.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "id": "1e99a5b8",
      "metadata": {},
      "source": [
        "# 🍎 Phi-4 Model with AIProjectClient 🍏\n",
        "\n",
        "**Phi-4** is a next-generation open model that aims to provide near GPT-4o capabilities at a fraction of the cost, making it ideal for many enterprise or personal use cases. It's especially great for chain-of-thought reasoning and RAG (Retrieval Augmented Generation) scenarios.\n",
        "\n",
        "In this notebook, you'll see how to:\n",
        "1. **Initialize** an `AIProjectClient` for your Azure AI Foundry environment.\n",
        "2. **Chat** with the **Phi-4** model using `azure-ai-inference`.\n",
        "3. **Show** a Health & Fitness example, featuring disclaimers and wellness Q&A.\n",
        "4. **Enjoy** the value proposition of a cheaper alternative to GPT-4 with strong reasoning capabilities. 🏋️\n",
        "\n",
        "> **Disclaimer**: This is not medical advice. Please consult professionals.\n",
        "\n",
        "## Why Phi-4?\n",
        "Phi-4 is a 14B-parameter model trained on curated data for high reasoning performance.\n",
        "- **Cost-Effective**: Get GPT-4-level performance for many tasks without the GPT-4 price.\n",
        "- **Reasoning & RAG**: Perfect for chain-of-thought reasoning steps and retrieval augmented generation workflows.\n",
        "- **Generous Context Window**: 16K tokens, enabling more context or longer user conversations.\n",
        "\n",
        "<img src=\"./seq-diagrams/4-phi-4.png\" width=\"30%\"/>\n"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "e93357dd",
      "metadata": {},
      "source": [
        "## 1. Setup\n",
        "\n",
        "Below, we'll install and import the necessary libraries:\n",
        "- **azure-ai-projects**: For the `AIProjectClient`.\n",
        "- **azure-ai-inference**: For calling your model, specifically the chat completions.\n",
        "- **azure-identity**: For `DefaultAzureCredential`.\n",
        "\n",
        "Ensure you have a `.env` file with:\n",
        "```bash\n",
        "PROJECT_CONNECTION_STRING=<your-conn-string>\n",
        "SERVERLESS_MODEL_NAME=phi-4\n",
        "```\n",
        "\n",
        "> **Note**: It's recommended to complete the [`3-basic-rag.ipynb`](./3-basic-rag.ipynb) notebook before this one, as it covers important concepts that will be helpful here."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "8b5634a0",
      "metadata": {},
      "outputs": [],
      "source": [
        "import os\n",
        "from dotenv import load_dotenv\n",
        "from pathlib import Path\n",
        "from azure.identity import DefaultAzureCredential\n",
        "from azure.ai.projects import AIProjectClient\n",
        "from azure.ai.inference.models import SystemMessage, UserMessage, AssistantMessage\n",
        "\n",
        "from pathlib import Path\n",
        "\n",
        "# Load environment variables\n",
        "notebook_path = Path().absolute()\n",
        "parent_dir = notebook_path.parent\n",
        "load_dotenv(parent_dir / '.env')\n",
        "\n",
        "conn_string = os.getenv(\"PROJECT_CONNECTION_STRING\")\n",
        "phi4_deployment = os.getenv(\"SERVERLESS_MODEL_NAME\", \"phi-4\")\n",
        "\n",
        "try:\n",
        "    project_client = AIProjectClient.from_connection_string(\n",
        "        credential=DefaultAzureCredential(),\n",
        "        conn_str=conn_string,\n",
        "    )\n",
        "    print(\"✅ AIProjectClient created successfully!\")\n",
        "except Exception as e:\n",
        "    print(\"❌ Error creating AIProjectClient:\", e)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "500d63ef",
      "metadata": {},
      "source": [
        "## 2. Chat with Phi-4 🍏\n",
        "We'll demonstrate a simple conversation using **Phi-4** in a health & fitness context. We'll define a system prompt that clarifies the role of the assistant. Then we'll ask some user queries.\n",
        "\n",
        "> Notice that Phi-4 is well-suited for chain-of-thought reasoning. We'll let it illustrate its reasoning steps for fun.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "0bcc4772",
      "metadata": {},
      "outputs": [],
      "source": [
        "def chat_with_phi4(user_question, chain_of_thought=False):\n",
        "    \"\"\"Send a chat request to the Phi-4 model with optional chain-of-thought.\"\"\"\n",
        "    # We'll define a system message with disclaimers:\n",
        "    system_prompt = (\n",
        "        \"You are a Phi-4 AI assistant, focusing on health and fitness.\\n\"\n",
        "        \"Remind users that you are not a medical professional, but can provide general info.\\n\"\n",
        "    )\n",
        "\n",
        "    # We can optionally instruct the model to show chain-of-thought. (Use carefully in production.)\n",
        "    if chain_of_thought:\n",
        "        system_prompt += \"Please show your step-by-step reasoning in your answer.\\n\"\n",
        "\n",
        "    # We create messages for system + user.\n",
        "    system_msg = SystemMessage(content=system_prompt)\n",
        "    user_msg = UserMessage(content=user_question)\n",
        "\n",
        "    with project_client.inference.get_chat_completions_client() as chat_client:\n",
        "        response = chat_client.complete(\n",
        "            model=phi4_deployment,\n",
        "            messages=[system_msg, user_msg],\n",
        "            temperature=0.8,  # a bit creative\n",
        "            top_p=0.9,\n",
        "            max_tokens=400,\n",
        "        )\n",
        "\n",
        "    return response.choices[0].message.content\n",
        "\n",
        "# Example usage:\n",
        "question = \"I'm training for a 5K. Any tips on a weekly workout schedule?\"\n",
        "answer = chat_with_phi4(question, chain_of_thought=True)\n",
        "print(\"🗣️ User:\", question)\n",
        "print(\"🤖 Phi-4:\", answer)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "bc68c40d",
      "metadata": {},
      "source": [
        "## 3. RAG-like Example (Stub)\n",
        "Phi-4 also excels in retrieval augmented generation scenarios, where you provide external context and let the model reason over it. Below is a **stub** example showing how you'd pass retrieved text as context.\n",
        "\n",
        "> In a real scenario, you'd embed & search for relevant passages, then feed them into the user/system message.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "419ea578",
      "metadata": {},
      "outputs": [],
      "source": [
        "def chat_with_phi4_rag(user_question, retrieved_doc):\n",
        "    \"\"\"Simulate an RAG flow by appending retrieved context to the system prompt.\"\"\"\n",
        "    system_prompt = (\n",
        "        \"You are Phi-4, helpful fitness AI.\\n\"\n",
        "        \"We have some context from the user's knowledge base: \\n\"\n",
        "        f\"{retrieved_doc}\\n\"\n",
        "        \"Please use this context to help your answer. If the context doesn't help, say so.\\n\"\n",
        "    )\n",
        "\n",
        "    system_msg = SystemMessage(content=system_prompt)\n",
        "    user_msg = UserMessage(content=user_question)\n",
        "\n",
        "    with project_client.inference.get_chat_completions_client() as chat_client:\n",
        "        response = chat_client.complete(\n",
        "            model=phi4_deployment,\n",
        "            messages=[system_msg, user_msg],\n",
        "            temperature=0.3,\n",
        "            max_tokens=300,\n",
        "        )\n",
        "    return response.choices[0].message.content\n",
        "\n",
        "# Let's define a dummy doc snippet:\n",
        "doc_snippet = \"Recommended to run 3 times per week and mix with cross-training.\\n\" \\\n",
        "              \"Include rest days or active recovery days for muscle repair.\"\n",
        "\n",
        "user_q = \"How often should I run weekly to prepare for a 5K?\"\n",
        "rag_answer = chat_with_phi4_rag(user_q, doc_snippet)\n",
        "print(\"🗣️ User:\", user_q)\n",
        "print(\"🤖 Phi-4 (RAG):\", rag_answer)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "3a33a375",
      "metadata": {},
      "source": [
        "## 4. Wrap-Up & Best Practices\n",
        "1. **Chain-of-Thought**: Great for debugging or certain QA tasks, but be mindful about revealing chain-of-thought to end users.\n",
        "2. **RAG**: Use `azure-ai-inference` with retrieval results to ground your answers.\n",
        "3. **OpenTelemetry**: Optionally integrate `opentelemetry-sdk` and `azure-core-tracing-opentelemetry` for full observability.\n",
        "4. **Evaluate**: Use `azure-ai-evaluation` to measure your model’s performance.\n",
        "5. **Cost & Performance**: Phi-4 aims to provide near GPT-4 performance at lower cost. Evaluate for your domain needs.\n",
        "\n",
        "## 🎉 Congratulations!\n",
        "You've seen how to:\n",
        "- Use **Phi-4** with `AIProjectClient` and `azure-ai-inference`.\n",
        "- Create a **chat** flow with chain-of-thought.\n",
        "- Stub a **RAG** scenario.\n",
        "\n",
        "> Happy hacking! 🏋️\n"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": ".venv",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.11.11"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 5
}