Tabtin

A self-hostable, mobile-first web app to create tables from images. It’s 2025, so of course this uses AI to help you create things like inventory lists and more!

Table of Contents

Overview 1.

Prerequisites 1.

Installation 1.

Getting Started 1.

AI Model Options

5.1 Google AI Studio (Free Tier)

5.2 OpenRouter (Paid/Free Tier)

5.3 Local AI Models

Development Setup 1.

Troubleshooting

1. Overview

Tabtin is an image processing platform that enables users to:

Define the data structure you want to extract from images (essentially create the first row of your future spreadsheet)

Select the vision model for data extraction (Google, OpenRouter, or local models via OpenAI-compatible endpoints)

Review extracted data in an intuitive interface (because who fully trusts AI?)

Export your data as CSV

The app handles the heavy lifting by sending batched images (e.g., front and back views of an object) to vision LLMs and extracting user-defined data. It includes review and redo methods to ensure you only capture the right data. Everything is customizable: prompts, field descriptions, context for the LLM, and your choice of model.

2. Prerequisites

Before starting, ensure you have:

Docker & Docker Compose (for containerized deployment)

Git (to clone the repository)

A Vision LLM (for example):

Google AI Studio API (free tier available)

OpenRouter API (free tier with 10€ credit)

Local LLM setup

See the AI Model Options section below for details on obtaining API keys and setting up models.

3. Installation

3.1 Clone the Repository

git  clone  https://github.com/janbndrf/tabtin.git

cd  tabtin

3.2 Configure Environment

Copy the example environment file:

cp  .env.example  .env

Edit .env and configure the following:

POCKETBASE_ADMIN_EMAIL=admin@example.com

POCKETBASE_ADMIN_PASSWORD=your-secure-admin-password

Note: These credentials secure your database backend. They are not used for user login.

3.3 Start with Docker

Start the application:

docker  compose  up  --build  -d

The first startup might take a couple of minutes. You can check the docker logs to ensure everything works as expected.

Access Points:

Frontend: http://YOUR-MACHINE-IP

PocketBase Admin: http://YOUR-MACHINE-IP:8090/_/ (database access - rarely needed)

4. Getting Started

4.1 Creating Your First Project

Each project results in a table. For example, imagine creating an inventory list of containers in your garage. This will be a project, where you will take images of each container in your garage.

1. Register & Login

Navigate to the register page and create your first account

Login to access the dashboard

1. Create a Project

Create a new project from the dashboard

Navigate to your project settings

1. Configure Model

Choose the model you want to use (Google, OpenRouter, or OpenAI-compatible endpoint for local models)

For cloud models with free tiers, set the max requests per minute to stay within rate limits

1. Define Table Structure

Set up your table structure (column names)

Add explanations for how you expect each field to look

This helps guide the AI in extracting the right data

1. Start Processing

Save your project settings

Begin taking or uploading images for processing

5. AI Model Options

Tabtin supports multiple vision models. You should definitely experiment to find models suitable for your use case and individual prompts.

Performance Characteristics:

Smarter models: Take more time per image but can extract more “abstract” data.

Smaller models: Faster but may struggle with complex tasks. They excel at extracting specific text or numbers present on images (e.g., cataloging containers)

Disclaimer: Currently, only Qwen3 model output produces reliable bounding boxes. Gemini output has been challenging to convert to proper bounding boxes, though Gemini’s extraction accuracy is excellent. If id have to accept unreviewed data, id probably go with gemini anytime.

5.1 Google AI Studio (Free Tier)

Google offers free access to their AI models through AI Studio with generous rate limits.

for example

gemini-2.5-flash-lite: Up to 1,500 requests per day

gemini-2.5-flash: 50-250 requests per day (still plenty for most use cases)

Get Your API Key:

• Visit: https://aistudio.google.com/api-keys

5.2 OpenRouter (Paid/Free Tier)

OpenRouter offers a unique free tier where you can use several models at no cost as long as you maintain a 10€ balance in your account.

How it Works:

Deposit 10€

Access multiple models from their free tier

Experiment with different models to find what works best

Get Your API Key:

• Visit: https://openrouter.ai

5.3 Local AI Models

For local vision LLMs, the Qwen3 series with llama.cpp is strongly recommended.

Important Notes:

All testing has been done with Qwen models

Qwen3 is currently the only model that works with 100% of this application’s functions

Other models, particularly for bounding box generation, have not been fully tested

Setup:

Configure your local endpoint as an OpenAI-compatible API

Be sure to add /v1/chat/completions endpoint!

6. Development Setup

6.1 Local Development

Install dependencies:

npm  install

Update .env for local development:

POCKETBASE_URL=http://127.0.0.1:8090

6.2 Running Development Servers

Open two terminal windows:

Terminal 1 - Backend:

npm  run  backend

Backend runs at http://localhost:8090

Terminal 2 - Frontend:

npm  run  dev

Frontend runs at http://localhost:5173

Both services are accessible from any device on your network, making it easy to test on mobile devices.

7. Troubleshooting

7.1 Poor Extraction Results

Try a different model: Smarter models often perform better

Improve your prompt: Add more specific context

Image quality: Ensure images are clear and well-lit

Field descriptions: Add detailed descriptions to guide the extraction

7.2 Bounding Boxes Not Working

Currently, only Qwen3-VL bounding boxes work reliably. Other models are not fully supported for this feature.

7.3 Rate Limiting Issues

If you’re hitting rate limits with cloud providers:

• Reduce the max requests per minute in project settings