🍎 Phi-4 Model with AIProjectClient 🍏¶

Phi-4 is a next-generation open model that aims to provide near GPT-4 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.

In this notebook, you'll see how to:

Initialize an AIProjectClient for your Azure AI Foundry environment.
Chat with the Phi-4 model using azure-ai-inference.
Show a Health & Fitness example, featuring disclaimers and wellness Q&A.
Enjoy the value proposition of a cheaper alternative to GPT-4 with strong reasoning capabilities. 🏋️

Disclaimer: This is not medical advice. Please consult professionals.

Why Phi-4?¶

Phi-4 is a 14B-parameter model trained on curated data for high reasoning performance.

Cost-Effective: Get GPT-4-level performance for many tasks without the GPT-4 price.
Reasoning & RAG: Perfect for chain-of-thought reasoning steps and retrieval augmented generation workflows.
Generous Context Window: 16K tokens, enabling more context or longer user conversations.

1. Setup¶

Below, we'll install and import the necessary libraries:

azure-ai-projects: For the AIProjectClient.
azure-ai-inference: For calling your model, specifically the chat completions.
azure-identity: For DefaultAzureCredential.

Ensure you have a .env file with:

PROJECT_CONNECTION_STRING=<your-conn-string>
SERVERLESS_MODEL_NAME=phi-4

Replace <your-conn-string> with your actual Azure AI Foundry project connection string.

In [ ]:

  Copied!     
 
import os
from dotenv import load_dotenv
from pathlib import Path
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from azure.ai.inference.models import SystemMessage, UserMessage, AssistantMessage

load_dotenv()

conn_string = os.getenv("PROJECT_CONNECTION_STRING")
phi4_deployment = os.getenv("SERVERLESS_MODEL_NAME", "phi-4")

try:
    project_client = AIProjectClient.from_connection_string(
        credential=DefaultAzureCredential(),
        conn_str=conn_string,
    )
    print("✅ AIProjectClient created successfully!")
except Exception as e:
    print("❌ Error creating AIProjectClient:", e)
import os from dotenv import load_dotenv from pathlib import Path from azure.identity import DefaultAzureCredential from azure.ai.projects import AIProjectClient from azure.ai.inference.models import SystemMessage, UserMessage, AssistantMessage load_dotenv() conn_string = os.getenv("PROJECT_CONNECTION_STRING") phi4_deployment = os.getenv("SERVERLESS_MODEL_NAME", "phi-4") try: project_client = AIProjectClient.from_connection_string( credential=DefaultAzureCredential(), conn_str=conn_string, ) print("✅ AIProjectClient created successfully!") except Exception as e: print("❌ Error creating AIProjectClient:", e)

2. Chat with Phi-4 🍏¶

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.

Notice that Phi-4 is well-suited for chain-of-thought reasoning. We'll let it illustrate its reasoning steps for fun.

In [ ]:

  Copied!     
 
def chat_with_phi4(user_question, chain_of_thought=False):
    """Send a chat request to the Phi-4 model with optional chain-of-thought."""
    # We'll define a system message with disclaimers:
    system_prompt = (
        "You are a Phi-4 AI assistant, focusing on health and fitness.\n"
        "Remind users that you are not a medical professional, but can provide general info.\n"
    )

    # We can optionally instruct the model to show chain-of-thought. (Use carefully in production.)
    if chain_of_thought:
        system_prompt += "Please show your step-by-step reasoning in your answer.\n"

    # We create messages for system + user.
    system_msg = SystemMessage(content=system_prompt)
    user_msg = UserMessage(content=user_question)

    with project_client.inference.get_chat_completions_client() as chat_client:
        response = chat_client.complete(
            model=phi4_deployment,
            messages=[system_msg, user_msg],
            temperature=0.8,  # a bit creative
            top_p=0.9,
            max_tokens=400,
        )

    return response.choices[0].message.content

# Example usage:
question = "I'm training for a 5K. Any tips on a weekly workout schedule?"
answer = chat_with_phi4(question, chain_of_thought=True)
print("🗣️ User:", question)
print("🤖 Phi-4:", answer)
def chat_with_phi4(user_question, chain_of_thought=False): """Send a chat request to the Phi-4 model with optional chain-of-thought.""" # We'll define a system message with disclaimers: system_prompt = ( "You are a Phi-4 AI assistant, focusing on health and fitness.\n" "Remind users that you are not a medical professional, but can provide general info.\n" ) # We can optionally instruct the model to show chain-of-thought. (Use carefully in production.) if chain_of_thought: system_prompt += "Please show your step-by-step reasoning in your answer.\n" # We create messages for system + user. system_msg = SystemMessage(content=system_prompt) user_msg = UserMessage(content=user_question) with project_client.inference.get_chat_completions_client() as chat_client: response = chat_client.complete( model=phi4_deployment, messages=[system_msg, user_msg], temperature=0.8, # a bit creative top_p=0.9, max_tokens=400, ) return response.choices[0].message.content # Example usage: question = "I'm training for a 5K. Any tips on a weekly workout schedule?" answer = chat_with_phi4(question, chain_of_thought=True) print("🗣️ User:", question) print("🤖 Phi-4:", answer)

3. RAG-like Example (Stub)¶

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.

In a real scenario, you'd embed & search for relevant passages, then feed them into the user/system message.

In [ ]:

  Copied!     
 
def chat_with_phi4_rag(user_question, retrieved_doc):
    """Simulate an RAG flow by appending retrieved context to the system prompt."""
    system_prompt = (
        "You are Phi-4, helpful fitness AI.\n"
        "We have some context from the user's knowledge base: \n"
        f"{retrieved_doc}\n"
        "Please use this context to help your answer. If the context doesn't help, say so.\n"
    )

    system_msg = SystemMessage(content=system_prompt)
    user_msg = UserMessage(content=user_question)

    with project_client.inference.get_chat_completions_client() as chat_client:
        response = chat_client.complete(
            model=phi4_deployment,
            messages=[system_msg, user_msg],
            temperature=0.3,
            max_tokens=300,
        )
    return response.choices[0].message.content

# Let's define a dummy doc snippet:
doc_snippet = "Recommended to run 3 times per week and mix with cross-training.\n" \
              "Include rest days or active recovery days for muscle repair."

user_q = "How often should I run weekly to prepare for a 5K?"
rag_answer = chat_with_phi4_rag(user_q, doc_snippet)
print("🗣️ User:", user_q)
print("🤖 Phi-4 (RAG):", rag_answer)
def chat_with_phi4_rag(user_question, retrieved_doc): """Simulate an RAG flow by appending retrieved context to the system prompt.""" system_prompt = ( "You are Phi-4, helpful fitness AI.\n" "We have some context from the user's knowledge base: \n" f"{retrieved_doc}\n" "Please use this context to help your answer. If the context doesn't help, say so.\n" ) system_msg = SystemMessage(content=system_prompt) user_msg = UserMessage(content=user_question) with project_client.inference.get_chat_completions_client() as chat_client: response = chat_client.complete( model=phi4_deployment, messages=[system_msg, user_msg], temperature=0.3, max_tokens=300, ) return response.choices[0].message.content # Let's define a dummy doc snippet: doc_snippet = "Recommended to run 3 times per week and mix with cross-training.\n" \ "Include rest days or active recovery days for muscle repair." user_q = "How often should I run weekly to prepare for a 5K?" rag_answer = chat_with_phi4_rag(user_q, doc_snippet) print("🗣️ User:", user_q) print("🤖 Phi-4 (RAG):", rag_answer)

4. Wrap-Up & Best Practices¶

Chain-of-Thought: Great for debugging or certain QA tasks, but be mindful about revealing chain-of-thought to end users.
RAG: Use azure-ai-inference with retrieval results to ground your answers.
OpenTelemetry: Optionally integrate opentelemetry-sdk and azure-core-tracing-opentelemetry for full observability.
Evaluate: Use azure-ai-evaluation to measure your model’s performance.
Cost & Performance: Phi-4 aims to provide near GPT-4 performance at lower cost. Evaluate for your domain needs.

🎉 Congratulations!¶

You've seen how to:

Use Phi-4 with AIProjectClient and azure-ai-inference.
Create a chat flow with chain-of-thought.
Stub a RAG scenario.

Happy hacking! 🏋️