Apppublic

olasogba/meridian-support

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

App README

Meridian Electronics — Customer Support Chatbot

A production-style customer support assistant for Meridian Electronics, built on top of an existing internal MCP server. The bot can:

—Browse and search the product catalog
—Authenticate returning customers via email + 4-digit PIN
—Show a logged-in customer their order history
—Place a new order with an explicit two-step confirmation

Architecture

   ┌─────────┐    HTTPS     ┌────────────────┐    Streamable HTTP     ┌─────────────┐
   │ Browser │ ───────────▶ │ Gradio + Agent │ ────────────────────▶ │ Order MCP   │
   │ (chat)  │              │  (this app)    │                        │  server     │
   └─────────┘              └────────────────┘                        └─────────────┘
                                    │
                                    │ uses
                                    ▼
                            OpenAI Responses API
                                (gpt-4o-mini)

—Agent — OpenAI Agents SDK, model gpt-4o-mini (cheap mini-tier).
—Tools — every MCP tool is wrapped in a Python @function_tool. The wrappers carry session state (Session dataclass) and inject the authenticated customer ID for privileged calls. The LLM never sees a customer_id parameter for orders or history; it cannot impersonate.
—Order placement is a two-phase flow: prepare_order re-fetches live prices and stages a pending order on the session; place_order takes no arguments and submits whatever was staged. The agent must obtain explicit customer confirmation between the two calls.
—MCP transport — Streamable HTTP. A fresh client session is opened per tool call (stateless, easy to reason about, latency is fine for chat).
—UI — Gradio chat. Session state lives in gr.State, isolated per browser session.

Files

File	Purpose
`app.py`	Gradio UI + agent run loop
`agent_setup.py`	Builds the Agent with the system prompt and tools
`tools.py`	Session-aware function tools that wrap MCP calls
`mcp_client.py`	Streamable-HTTP client wrapper
`prompts.py`	System prompt for "Meri"
`config.py`	Env-driven configuration
`discover.py`	Standalone script to dump the MCP server's tool schema
`_smoke_test.py`	End-to-end agent test (no UI)

Run locally

bash

uv sync
cp .env.example .env   # then edit and set OPENAI_API_KEY
uv run python app.py

Open http://127.0.0.1:7860.

Environment

—OPENAI_API_KEY — required.
—MODEL — defaults to gpt-4o-mini.
—MCP_SERVER_URL — defaults to the Meridian MCP server.

On Hugging Face Spaces, set OPENAI_API_KEY as a Space Secret (Settings → Variables and secrets → New secret).

Security notes

—Authentication is by email + 4-digit PIN against the MCP server.
—The bot only ever calls privileged tools with the customer ID it captured from a successful PIN verification. UUIDs typed into the chat are ignored.
—Order details are gated: my_order_details checks the order belongs to the authenticated customer before fetching it.
—PINs and full API keys are never logged.

Limitations / future work

—No persistent conversation history across sessions.
—Price extraction from MCP get_product text uses a regex; a structured response from the MCP server would be safer.
—No rate limiting on the chat endpoint.
—Single-region MCP server; a regional outage would take the bot down.
—gpt-4o-mini will occasionally over-confirm or hallucinate a SKU; tightening the system prompt or moving to gpt-4.1-mini would help.