UnderDoc Tutorial - Expense Analytics using GenAI and MCP server for SQLite DB

- UnderDoc Tutorial - Expense Analytics using GenAI and MCP server for SQLite DB
- Introduction
- Overview
- What is MCP
- Prerequisite
- Get an UnderDoc API Key
- Step 1: Signup
- Step 2: Login
- Step 3: Get the key from "Key Management" Page
- Install uv
- Install Python 3.12
- Install sqlite command line interface
- Environment Setup
- Clone the project for this tutorial
- Setup Python Virtual Environment and Install Required Packages
- Environment Verification
- Initialize a SQLite DB (optional)
- Copy your expense images into the folder receipt-images
- Perform Data Extraction and Save to DB
- Data Verification
- Expense Analytics with Natural Language
- Install Claude Desktop
- Clone the source of reference MCP server from Github
- Install MCP Server for SQLite for Claude Desktop
- Verify the MCP server in Claude Desktop
- Overview of UnderDoc Expense (SQLite) DB
- Expense Analytics by Asking Questions
- Final words
- Resources

---

Introduction

In my previous tutorial, I discussed about how to extract structure expense data from images, store it into a SQLite DB and use Metabase to perform expense analytics.

In this tutorial, I would like to discuss using GenAI and MCP server for SQLite in performing expense analytics by chatting with a LLM in natural language.

This tutorial focus on using GenAI (Claude Desktop) and MCP server for SQLite, and assume that the database was already populated with UnderDoc expense records. You can follow the previous tutorial up to the section "Perform Data Extraction and Save to DB" (skip the rest of the sections relating to Metabase). If you want to have a quick look on how the MCP server works, you can also download the sample SQLite DB with 20 sample expense records here.

You need to have the following in order to be able to run through this tutorial:

- A laptop or desktop running MacOS
- uv (a Python runtime and package manager for running the SQLite MCP server)
- git (for cloning the MCP server source code from github)
- Claude Desktop (this is the GenAI app we are using in this tutorial, which supports MCP servers)
- Sqlite DB

For those reader haven't went through the previous tutorial, I have replicated the steps here for easy reference. If you already have the SQLite DB ready, you can skip to the section "Expense Analytics with Natural Language".

Overview

This tutorial use the following components for various tasks:

- UnderDoc for extracting data from receipt/invoice/demand-notes images
- uv for managing Python runtime and runs the MCP servers
- SQLite for storing extracted data
- Claude Desktop for Mac for expense analytics using LLM with MCP servers

What is MCP

There are too many articles discussing MCP (Model Context Protocol), so I will not go into details here. In short, MCP provided a standardized way for LLM to interact with various resources and tools required for answering users' questions.

There are already a numbers of tools supports MCP server, like Claude desktop, Cursor IDE, etc. Many other GenAI providers (e.g. OpenAI, Google, etc.) also announced the support of MCP in their services as well.

The open source MCP repo in Github had provided a list of reference servers. There is one for SQLite DB integration, which will be used in this tutorial.

Prerequisite

Get an UnderDoc API Key

The key step in this tutorial is to extract the expense information from images into structured data (e.g. shop name, amount, currency, inferred category, etc.) that is suitable for storing into a database for exploration. We will use UnderDoc, which is a cost-effective platform (with a free tier) in performing this task. This tool also works well with multi-language images.

Using UnderDoc is simple, just signup at their developer portal and get an API key.

Step 1: Signup

First of all, visit the signup page at UnderDoc Developer Portal

Enter your email, password and then confirm Sign Up.

Upon sign up successful, you will receive an email. You need to verify your email before login.

Step 2: Login

Once sign up success and email verified, you can then login to UnderDoc Developer Portal to login.

Step 3: Get the key from "Key Management" Page

On the left, select "Key Management".

In the Key Management page, click the eye icon to display the key, and then copy it for use later in this tutorial.

Now you have the UnderDoc API key, and can proceed to the rest of the tutorial.

Install uv

To run the SQLite MCP server, "uv" is required. You can refer to their site for installation instructions. The easiest way is to install it using brew:

brew install uv

Install Python 3.12

This tutorial was tested with Python 3.12. To install this into your Mac, the easiest way is to install via homebrew. The following is the command.

brew install python@3.12

Install sqlite command line interface

The expense data extracted will be persisted to SQLite DB locally. We will use SQLite to verify the database as well as the data stored. To install, we also use homebrew.

brew install sqlite

This concludes all prerequisites, let's proceed to environment setup!

Environment Setup

Clone the project for this tutorial

All the Python scripts required, and some sample images were provided by this tutorial. Clone it from GitHub with the following command.

git clone https://github.com/under-doc/underdoc-expense-analytics-tutorial.git

Once cloned, change directory into the project folder. For example, my folder is as follows:

/Users/clarenceho/workspace/underdoc/underdoc-expense-analytics-tutorial

In this folder, you will see some folders and files. The following is an overview:

- README.md: this file
- model.py: the Python script for model classes
- main.py: the Python script for the extraction process
- requirements.txt: the Python library packages that this project requires
- receipt-images: the expense images should be put here. You can see 20 sample images provided
- metabase-data: this is where the Metabase and UnderDoc database will be stored

Setup Python Virtual Environment and Install Required Packages

The next step is to create a Python virtual environment for this tutorial and install the required packages. Run the following commands in order.

```bash