Azure foundry file search - setup and deploying as a remote agent (not ephemeral agent)

 File search allow us to feed information into our model 

In this implementation we are deploying agent in Azure Foundry and it uses chatgpt-5-mini. I created the file search manually. You agent can only have 1 index.  File search is not RAG. 

With RAG you use Azure AI Search. File search is pretty generic. It requires basic setup where you don't necessary have to create a storage account manually to host your file. 

As you can see here, I am uploading zippolock product info and it automatically create an index and embedd the info for me. There are limits to the file that you will be uploading. 

And then I save this as an agent as shown here :-

Ensuring we have the right pypi dependencies

dependencies = [

     "agent-framework>=1.13.0",

     "azure-ai-projects", # Main Azure Foundry SDK

     "azure-identity", # For authentication

]

And then we can use the following code to query what is zippolock. The magic is to INSTRUCT the agent to call your filesearch toolbox with this command here :- "use myfilesearch toolbox and

find out what is zippolock?"

import asyncio

import httpx

from azure.identity import DefaultAzureCredential, get_bearer_token_provider

from agent_framework import MCPStreamableHTTPTool

from agent_framework_foundry import FoundryChatClient

# ── Configuration ─────────────────────────────────────────────────────────────

endpoint = "https://your-foundry-instance.services.ai.azure.com/api/projects/proj-default"

toolbox_name = "myfilesearch"

toolbox_version = "1"

model_deployment = "gpt-5.4-mini"

from urllib.parse import urlparse

_parsed = urlparse(endpoint)

toolbox_url = f"{endpoint.rstrip('/')}/toolboxes/{toolbox_name}/versions/{toolbox_version}/mcp?api-version=v1"

# ── Reusable functions (can be pulled into a hosted agent main.py) ────────────

# Toolbox MCP auth

class _ToolboxAuth(httpx.Auth):

    """Injects a fresh bearer token on every request."""

    def __init__(self, token_provider):

        self._get_token = token_provider

    def auth_flow(self, request):

        request.headers["Authorization"] = f"Bearer {self._get_token()}"

        yield request

# [START msft_agentframework_toolbox]

_agent = None

_toolbox = None

async def create_agent_with_toolbox():

    """Create an Agent Framework agent wired to a Foundry toolbox via MCP."""

    global _agent, _toolbox

    credential = DefaultAzureCredential()

    token_provider = get_bearer_token_provider(

        credential, "https://ai.azure.com/.default"

    )

    http_client = httpx.AsyncClient(

        auth=_ToolboxAuth(token_provider),

        headers={"Foundry-Features": "Toolboxes=V1Preview"},

        timeout=120.0,

    )

    _toolbox = MCPStreamableHTTPTool(

        name=toolbox_name,

        url=toolbox_url,

        http_client=http_client,

        load_prompts=False,

    )

    chat_client = FoundryChatClient(

        project_endpoint=endpoint,

        model=model_deployment,

        credential=credential,

    )

    _agent = chat_client.as_agent(

        name="toolbox-agent",

        instructions="You are a helpful assistant with access to Azure AI Foundry toolbox tools.",

        tools=[_toolbox],

    )

async def call_agent_with_toolbox(user_input: str):

    """Send a message to the toolbox agent and print the response."""

    response = await _agent.run(messages=user_input, stream=False)

    print(response.text)

async def close_agent():

    """Close the toolbox MCP connection cleanly."""

    if _toolbox:

        await _toolbox.close()

# [END msft_agentframework_toolbox]

# ── Script entry point ────────────────────────────────────────────────────

async def main():

    await create_agent_with_toolbox()

    try:

        await call_agent_with_toolbox("use myfilesearch toolbox and

find out what is zippolock?")

    finally:

        await close_agent()

asyncio.run(main())

And the output is :-