LLM Extractinator

Turn unstructured text into Pydantic-validated structured data using local LLMs via Ollama.

⚠️ Prototype: this project is under active development. Interfaces, task formats, and directory names may change. Always inspect and validate the generated output before using it downstream.

LLM Extractinator lets you turn unstructured text (reports, notes, CSV text columns) into structured, Pydantic-validated data using local LLMs (via Ollama). It comes with:

• a Studio (Streamlit app) for no-code configuration
• a CLI (extractinate) for repeatable runs
• a parser builder (build-parser) to generate Pydantic models
• task JSON files to describe what to extract and from where

📚 Docs: https://diagnijmegen.github.io/llm_extractinator/

---

1. Prerequisites

This project expects a local LLM endpoint, currently Ollama.

Install Ollama

Linux:

curl -fsSL https://ollama.com/install.sh | sh

Windows / macOS: download from https://ollama.com/download

Make sure the Ollama service is running before you try to extract.

---

2. Installation

We recommend using a fresh environment:

conda create -n llm_extractinator python=3.11
conda activate llm_extractinator

Install from PyPI:

pip install llm_extractinator

Or install from source:

git clone https://github.com/DIAGNijmegen/llm_extractinator.git
cd llm_extractinator
pip install -e .

---

3. Quick start

A. Run with Docker (recommended for GPU systems)

You can run LLM Extractinator entirely via Docker, which includes Python, Ollama, and the Streamlit app in one container.

Make sure Docker is installed.
For GPU acceleration, also install the NVIDIA Container Toolkit.

Create local directories that will be mounted inside the container:

mkdir -p data examples tasks output ollama_models

Windows / PowerShell:

# Remove `--gpus all` if you don't have a GPU
docker run --rm --gpus all `
  -p 127.0.0.1:8501:8501 `
  -p 11434:11434 `
  -v ${PWD}/data:/app/data `
  -v ${PWD}/examples:/app/examples `
  -v ${PWD}/tasks:/app/tasks `
  -v ${PWD}/output:/app/output `
  -v ${PWD}/ollama_models:/root/.ollama `
  lmmasters/llm_extractinator:latest

Linux / macOS:

docker run --rm --gpus all \
  -p 127.0.0.1:8501:8501 \
  -p 11434:11434 \
  -v $(pwd)/data:/app/data \
  -v $(pwd)/examples:/app/examples \
  -v $(pwd)/tasks:/app/tasks \
  -v $(pwd)/output:/app/output \
  -v $(pwd)/ollama_models:/root/.ollama \
  lmmasters/llm_extractinator:latest

Note: The -v ${PWD}/ollama_models:/root/.ollama mount persists Ollama models between container runs, so you don't need to re-pull models each time. You can omit this mount if you don't mind re-downloading models.

This launches the Streamlit Studio on http://127.0.0.1:8501.
To open an interactive shell instead of the app, append shell to the command.

See docs/docker.md for full details.

---

B. Studio (local install)

Run the interactive UI:

launch-extractinator

This starts the Streamlit-based Studio (usually at http://localhost:8501) where you can:

• create/select datasets
• design a parser/output model
• create task JSON files
• run tasks and view logs

Anything you configure here can also be run from the CLI.

---

C. CLI

Run a task by ID:

extractinate --task_id 1 --model_name "phi4"

Common options:

• --task_dir tasks/ — where your task JSON files live
• --data_dir data/ — where your CSV/JSON input lives
• --output_dir output/ — where to write extracted results
• --run_name my_first_run — for easier tracking

See docs/cli.md for the full reference.

---

D. Python

You can also call it from Python:

from llm_extractinator import extractinate

extractinate(
    task_id=1,
    model_name="phi4",
    output_dir="output/",
)

---

4. Tasks

Tasks describe what to extract and from where. By convention they live in tasks/ and are named:

Task001_products.json
Task002_reports.json
...

A minimal task might look like:

{
  "Description": "Extract product data from CSV",
  "Data_Path": "products.csv",
  "Input_Field": "text",
  "Parser_Format": "product_parser.py"
}

• Data_Path: relative to your data directory
• Input_Field: column/key that contains the text
• Parser_Format: Python file in tasks/parsers/ that defines the Pydantic model

You can author these in Studio or by hand.

---

5. Parser builder

If you don’t want to write Pydantic models by hand:

build-parser

Export the generated model to:

tasks/parsers/<name>.py

Then reference that filename in the task JSON.

---

6. Documentation

See the docs/ directory (or the published site) for:

• data preparation
• parser UI
• CLI flags and examples
• Studio walkthrough
• manual/advanced running

---

7. Contributing

Pull requests are welcome. Please keep the CLI and the docs in sync (especially task naming and required fields).

---

8. Citation & Attribution

If you use this tool in your research, please cite:

https://doi.org/10.1093/jamiaopen/ooaf109

This tool was developed by the Oncology Research Group at the Diagnostic Image Analysis Group (DIAG), Radboud University Medical Center. 🔗 diagnijmegen.nl/research/oncology

Contact:

Name	Email
Luc Builtjes	[email protected]
Alessa Hering	[email protected]