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.

What This Is

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:

1. Calls search_metrics — retrieves the Return Rate definition, learns cancelled orders must be excluded from both numerator and denominator
2. Calls get_schema — discovers actual table and column names
3. Calls run_sql with wrong case — gets zeros, self-corrects by checking distinct status values
4. Calls run_sql again with correct values — returns accurate rates per segment

No framework orchestrated that. The agent reasoned through it.

Tools

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 Metrics Glossary

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

Dataset

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.

Quickstart

1. Clone and install

git clone https://github.com/sourabhsurana06/retail-analytics-agent cd retail-analytics-agent pip install -r requirements.txt

2. Set up environment

cp .env.example .env

Add your OPENAI_API_KEY

3. Build the database and vector index

python3 core/CreateDB.py python3 build_index.py

4. Start the MCP server

python3 server.py

5. Run the agent (second terminal)

python3 agent.py

Project Structure

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

What This Is Not

• 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.

License

MIT