ReportMiner 0.1.3

Easy to use wrapper for text document question and answer using RAG and LLM

ReportMiner is an easy-to-use wrapper for multimodal retrieval augmented generation (RAG) tasks for technical reports. With ReportMiner, you can use any HuggingFace or Byaldi and apply for your reports.

🤖 Installation

Make sure to install Poppler first

sudo apt-get install -y poppler-utils on Linux OS

Then, install the package

pip install reportminer

🚀 Use RAG in 3 simple steps

Using ReportMiner is just few lines of codes.

1. Setup embedding model and visual language model

Here we will use ColPali-1.2 as our embedding model and SmolVLM as our visual language model (by default). This combination works very well even with free-tier Colab GPU T4 and low memory.

from reportminer import rag

# Setup embedding and VL model
rag_models = rag.setup_model()

If you have access to Colab Pro's GPU A100 and high memory, you can make advantage of Qwen2-VL as your visual language model and activate Flash Attention 2.

2. Upload your report

# Specify pdf file
pdf_file = '/content/15-9-19a-core.pdf'

# Process PDF report by converting into embeddings
rag_models = rag.process_PDF(pdf_file, rag_models, dpi=200, index_name='pvt-rag')

3. Perform query

There are 2 minimal tasks you can use with ReportMiner. First, you can retrieve relevant page taking advantage of the embedding model

rag.RAG('How is the Klinkenberg corrected gas horizontal permeability to porosity looks like at core 7?', 
        rag_models, k=1)

Second, you can ask a question, taking advantage of the visual language model

rag.Ask('How many cored intervals are there in the report? Mention the depth of each', rag_models)

👨‍💻 Future development

In the next version, the following will be implemented:

1. Support for multiple PDF reports in a directory structure (best industrial case)
2. Support for fine tuning with user reports of specific domain
3. Integration with vector databases for robust document indexing