Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Retail Analytics Agent What is the return rate for each customer segment?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A from-scratch MCP (Model Context Protocol) server combining SQL query tools and RAG (Retrieval-Augmented Generation) into a single ReAct agent. Built without LangChain, LlamaIndex, or any agentic framework — just Python, Flask, FAISS, and the OpenAI API.
Most agentic AI tutorials either wrap everything in LangChain and hide what's actually happening, or demo a single tool (SQL or RAG) in isolation.
This project does neither. It builds a multi-tool MCP server where a reasoning agent decides in real time whether a question requires structured data retrieval (SQL), unstructured knowledge lookup (RAG), or both in sequence.
Example:
"What is the return rate for each customer segment? Use the correct metric definition."
The agent:
Calls search_metrics — retrieves the Return Rate definition, learns cancelled orders must be excluded from both numerator and denominator
Calls get_schema — discovers actual table and column names
Calls run_sql with wrong case — gets zeros, self-corrects by checking distinct status values
Calls run_sql again with correct values — returns accurate rates per segment
No framework orchestrated that. The agent reasoned through it.
Related MCP server: Modular RAG System
Tool | Type | Description |
get_schema | SQL | Returns all table names, column names, and data types |
run_sql | SQL | Executes a SELECT query, returns rows as JSON |
list_metrics | RAG | Returns all metric names and one-line descriptions |
search_metrics | RAG | Semantic search over the metrics glossary PDF |
The RAG knowledge base is a PDF containing precise business metric definitions with inclusion/exclusion rules. These are the distinctions a naive agent would get wrong without it:
Return Rate: cancelled orders excluded from both numerator and denominator
LTV Gross: returned orders included — this is a demand-side metric
LTV Net: returned orders netted to zero — this is the revenue-side metric
Category Affinity: returned items excluded — a return signals category rejection
Recent Purchase Activity: returned orders included — engagement, not revenue
Synthetic Indian retail database:
15 customers across 6 cities, segmented into Premium / Standard / Budget
15 products across 8 categories with rupee-denominated pricing
90 orders across 2024 with statuses: Completed / Returned / Pending
222 line items with quantity and discount percentage
Seeded deterministically (random.seed(42)) — results are reproducible.
git clone https://github.com/sourabhsurana06/retail-analytics-agent cd retail-analytics-agent pip install -r requirements.txt
cp .env.example .env
python3 core/CreateDB.py python3 build_index.py
python3 server.py
python3 agent.py
retail-analytics-agent/ ├── core/ │ ├── init.py │ ├── database.py │ ├── CreateDB.py │ ├── sql_tools.py │ ├── rag_tools.py │ └── retail_analytics_metric_list.pdf ├── data/ (gitignored — generated files) ├── agent.py ├── server.py ├── build_index.py ├── requirements.txt ├── .env.example └── .gitignore
Not production-ready (SQLite, no auth, single-threaded Flask dev server)
Not a framework demo — no LangChain, no LlamaIndex, no AutoGen
Not complete (no streaming, no async, no retry logic)
It is a learning system that shows exactly what is happening at each step.
MIT
This server cannot be deployed
Analytical memory for AI agents: a real Postgres queried in plain English over MCP. One command.
MCP server for building and testing AI agents with multi-model experimentation and insights.
MCP server unifying ERPs, CRMs, APIs and knowledge base for Claude, ChatGPT and Gemini.
Text-to-SQL MCP server: read-only queries on PostgreSQL, MySQL and SQL Server with your real schema