Vision LLM - Batch Document VQA with structured responses

Note

See details in the associated wiki article.

This repository uses Large Language Models with vision capabilities to extract information from collections of documents. The goal is to create a fully local pipeline that runs on a single machine, and can be used to extract information from document collections for usage in downstream tasks.

How to use

Installation

Run the following commands to clone the repository:

git clone https://github.com/IonMich/batch-doc-vqa.git
cd batch-doc-vqa

Create a new (conda) environment and install the required packages:

conda create -n batch python=3.11
conda activate batch
pip install -r requirements.txt

Prepare the documents

It is common for a batch of documents to be stored in a single large PDF file. Since multimodal LLMs can only process images, the first step is to convert the PDF file into a collection of images. (Of course, if you already have a collection of images, you can skip this step.) This can be done using the pdf_to_imgs.py script. For example, to convert a batch of 4-page documents into images stored in a single PDF file imgs/quiz11-presidents.pdf to images at 300 DPI, you can run the following command:

python pdf_to_imgs.py --filepath imgs/quiz11-presidents.pdf --pages_i 4 --dpi 300 --output_dir imgs/q11/

This will create a directory imgs/q11/ containing the images of the 4-page documents. Set the --pages_i argument to the number of pages in each document. The images will follow the naming convention doc-0-page-1-*.png, ..., doc-0-page-4-*.png, doc-1-page-1-*.png, ..., etc. A CSV file imgs/q11/doc_info.csv will also be created, containing the metadata of the images (the page number and the document number).

Example Usage

The outlines_quiz.py script can be used to extract information from a collection of documents using local Vision LLMs. For example, to extract the 8-digit university IDs from the documents in the imgs/q11/ directory, you can run the following command:

    python outlines_quiz.py

By default, this pipeline will use the Qwen2-VL-2B-Instruct model to extract full names, 8-digit university IDs, and section numbers from the documents. You can change the model used by setting the --model argument. You can change the information extracted by modifying the Pydantic schema in outlines_quiz.py. The extracted information will be saved in a JSON file in the tests/output/ directory.

Post-processing

You can analyze the extracted information using string matching algorithms. For example, take a look at the stringmatching.ipynb notebook to see how to match the extracted last names and university IDs to a list of entries in a CSV file.

Benchmarks

Our small test dataset (./imgs/quiz11-presidents.pdf) consists of 32 documents representing Physics quizzes and the task is to match them to the test students who took the quiz via their 8-digit university ID and, optionally, their names (./tests/data/test_ids.csv). We have already saturated our test dataset with 100% statistically confident detections, but more optimizations are explored to decrease inference cost. You can find more details here. Currently the best performing pipeline is one that uses outlines to enforce JSON schemas on the model's responses, and the Qwen2-VL series of models. It uses less than 5GB of VRAM and completes in about 2 minutes on an RTX 3060 Ti. See the code here. The pipeline has been tested only on Ubuntu 22.04 with an RTX 3060 Ti and 8GB of VRAM.

	OpenCV+CNN	outlines + VLM	outlines + VLM	VLM w/ structured output	VLM w/ structured output
LLM model	N/A	outlines + Qwen2-VL-2B-Instruct	outlines + SmolVLM	gpt-4o-mini	gemini-2.0-flash-exp
LLM model size	N/A	2B	2B	??	??
Open-weights	N/A	Yes (Apache 2.0 License)	Yes (Apache 2.0 License)	No	No
# samples	1	1	1	1	1
logits available	Yes	No	No	Yes (unused)	Yes (unused)
regex pattern	N/A	Yes	Yes	No	No
digit_top1	85.16%	98.44%	68.35%	87.89%	99.22%
digit_top2	90.62%	N/A	N/A	N/A	N/A
digit_top3	94.14%	N/A	N/A	N/A	N/A
8-digit id_top1	??	90.63%	53.13%	50.00%	93.75%
lastname_top1	N/A	100%	93.75%	93.75%	93.75%
Detect Type	ID (1)	LastName (2) + ID (1)	LastName (2) + ID (1)	LastName (2) + ID (1)	LastName (2) + ID (1)
ID Avg $d_\mathrm{Lev}$	N/A	0.1250	2.3750	0.6563	0.0625
Lastname Avg $d_\mathrm{Lev}$	N/A	0.0000	0.0938	0.156250	0.0625
Docs detected	90.62% (29/32)	100.00% (32/32)	68.75% (22/32)	100% (32/32)	100% (32/32)
Runtime	~ 1 second	~ 2 minutes (RTX 3060 Ti 8GB)	~ 2 minutes (RTX 3060 Ti 8GB)	~5 minutes (sequential)	~5 minutes (sequential)

[OLD] Ollama + Llama3.2-Vision 11B

Ollama Installation

First, install Ollama via the official installation instructions.
Pull the Llama3.2-Vision 11B model (or another model of your choice) on the Ollama server:
```
ollama pull llama3.2-vision
```
Install the Python Ollama wrapper. You can do this via pip in a dedicated conda environment:
```
conda create -n ollama python=3.11
conda activate ollama
pip install ollama
```

Clone this repository. There are no additional dependencies.

git clone https://github.com/IonMich/batch-doc-vqa
cd batch-doc-vqa

Start the Ollama server. On MacOS this can be done by simply opening the Ollama app. On Linux run ollama server in the terminal.
Check the SYSTEM_MESSAGES variable in llamavision.py. This are the prompts that the model will use to generate its responses (by default the prompt at index 0). Leave the prompt unchanged if you want to check the installation on the test images. Feel free to add your own images to the imgs directory, or to change the --filepath command line argument to point to a different directory.

That's it! You're ready to use the pipeline:

python llamavision.py --filepath imgs/sub-page-3.png --n_trials 10

Usage

The Ollama pipeline is a Python scriptwith the following command line usage:

    python llamavision.py [-h] [--filepath FILEPATH] [--pattern PATTERN] [--n_trials N_TRIALS] [--system SYSTEM] [--model MODEL] [--no-stream] [--top_k TOP_K]

The pipeline has the following options:

--filepath: (str) The path to the directory containing the images to be processed. If the path is a PNG file, the pipeline will process that single image. If the path is a directory, the pipeline will process (by default) all images in that directory.
--pattern: (str) A string that is used to filter the images in the directory. E.g. --pattern "page-3" will only process PNG files that contain the string page-3 in their filename. By default, this is set to "", which means that all images in the directory will be processed. This option is ignored if the --filepath is a single PNG file.
--n_trials: (int) The number of trials to run for each image. This can be useful to do multi-shot inference on the same image, or to get a sense of the variance in the model's predictions.
--system: (str) The system prompt index to use, which selects the prompt from a list of system prompts. By default, this is set to 0. See SYSTEM_PROMPTS in llamavision.py for the list of some example system prompts, and replace these with your own prompts.
--model: (str) The model to use. By default, this is set to llama3.2-vision, but any LLM with vision capabilities can be used.
--top_k: (int) The number of top answers to return. By default, this is set to 10. Lowering this number decreases creativity and variance in the model's predictions. Note that this is less than the Ollama default of 40.
--no-stream: (flag) If set, the pipeline will not stream the LLM responses in chunks. Instead, the pipeline will wait for the model to finish processing the image before printing the responses. Default (when this flag is omitted) is to stream the responses.

Motivations

Recent advances in LLM modelling have made it conceivable to build a quantifiably reliable pipeline to extract information in bulk from documents:

License

IonMich/batch-doc-vqa

ErrorLooks like something went wrong!

About

Topics

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Error
Looks like something went wrong!

Packages