Phi4 Agent demo for AIPC

supplementary_materials/phi4-agent/README.md

@@ -0,0 +1,48 @@
+# Phi-4-mini Agent on AI-PC with OpenVINO™ & 🤗 smolagents
+In this notebook we will guide you on how to deploy a [Phi-4-mini](https://huggingface.co/microsoft/Phi-4-mini-instruct) agent equipped with tools and MCP to run fully locally on your Intel<sup>®</sup> Core™ Ultra laptop with [OpenVINO™](https://docs.openvino.ai/2025/index.html) and [🤗 smolagents](https://github.com/huggingface/smolagents).
+
+An LLM based agent is a program where the LLM control the entire workflow of the program. In this notebook we build a multi-step agent. Given a task, the LLM will decide which action to take in each step and when to terminate and return a final answer to the user.
+
+We will demonstrate how the agent with the help of a few simple tools, like web-search and a code-completion tool (Also based on Phi-4-mini), the agent is able to code new tools and then use those tools summarize information into a PPT presentation for example.
+
+Furthermore, we will demonstrate how with the use of a YouTube transcript MCP we can chat with YouTube videos without worrying about the prompt exploding to thousands of tokens and everything is done locally with speed.
+
+## Prerequisites
+Create a new Python environment for this notebook and activate it and make sure you have `git` installed. E.g.
+```cmd
+conda create -n phi4-demo python=3.11
+conda activate phi4-demo
+```
+
+Then, lets start by installing all the packages we will need
+```cmd
+pip install -r requirements.txt
+```
+
+For this notebook we used a modified version of smolagents so we will need to clone the repository, apply our patch and install from source.
+```
+git clone https://github.com/huggingface/smolagents && cd smolagents
+git checkout v1.14.0 && git apply ..\phi4_smolagents.patch
+pip install .[mcp]
+cd ..
+```
+
+> [!NOTE]
+> To have the most up-to-date experience with OpenVINO you can also install the nightly versions:
+> ```cmd
+> pip install --pre -U openvino --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
+> pip install git+https://github.com/huggingface/optimum.git
+> pip install git+https://github.com/huggingface/optimum-intel.git
+> ```
+
+## Prepare Phi-4-mini model for inference
+We will use `optimum-cli` to convert and quantize Phi-4-mini model from the HuggingFace Hub to OpenVINO format
+```cmd
+optimum-cli export openvino --model microsoft/Phi-4-mini-instruct --task text-generation-with-past --weight-format int8 phi-4-mini-instruct-int8-ov
+```
+
+## Demo
+Now we are ready to run our demo:
+```cmd
+python phi4_agent.py
+```

supplementary_materials/phi4-agent/phi4_agent.py

@@ -0,0 +1,62 @@
+import argparse
+
+from tools import CodeCompletionTool, YoutubeTranscriptRetriever, build_presentation
+from mcp import StdioServerParameters
+
+from smolagents import TransformersModel, ToolCallingAgent, GradioUI
+from smolagents.mcp_client import MCPClient
+
+
+def add_arguments(parser):
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        default="phi-4-mini-instruct-int8-ov",
+        help="Path to the openvino model directory.",
+    )
+    parser.add_argument(
+        "--backend",
+        type=str,
+        default="ov-optimum",
+        choices=["ov-optimum", "torch"],
+        help="Backend to use for the model.",
+    )
+    return parser
+
+
+def main(model_path, backend="ov-optimum"):
+    # Initialize the Phi-4-mini model
+    model = TransformersModel(
+        model_id=model_path, max_new_tokens=1024, backend=backend)
+
+
+    server_params = StdioServerParameters(
+        command="python",
+        args=["-m", "duckduckgo_mcp_server.server"]
+    )
+    tools = [build_presentation, CodeCompletionTool(model.model, model.tokenizer, max_new_tokens=1024)]
+
+    # We initialize the MCP in a try block to ensure we disconnect from the MCP servers if there's an error
+    try:
+        # Start the MCP server
+        mcp_client = MCPClient(server_params)
+        tools.extend(mcp_client.get_tools())
+        # Youtube tool also uses an MCP server
+        yt_transcript_retriever = YoutubeTranscriptRetriever(device='GPU')
+        tools.append(yt_transcript_retriever)
+
+        # After initializing the tools, we can initialize our agent
+        agent = ToolCallingAgent(tools=tools, model=model, add_base_tools=False, max_steps=5)
+
+        # Now we can start our gradio demo, you can also run a single task with agent via agent.run()
+        GradioUI(agent).launch(inbrowser=True, inline=False)
+    finally:
+        mcp_client.disconnect()
+        yt_transcript_retriever.disconnect()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the Phi-4 agent.")
+    parser = add_arguments(parser)
+    args = parser.parse_args()
+    main(**vars(args))