PRAG: Parameterized Retrieval Augmentation Generation Tool for Improving the Performance of Q&A Systems

General Introduction

PRAG (Parametric Retrieval-Augmented Generation) is an innovative retrieval-augmented generation tool designed to enhance generation by embedding external knowledge directly into the parameter space of a Large Language Model (LLM). The tool overcomes the limitations of traditional contextual retrieval-augmented generation methods, reduces computational overhead, and enhances the model's reasoning and synthesis capabilities by deeply integrating external knowledge.PRAG provides end-to-end implementations including a data enhancement module, a parameter training module, and an inference module for performance testing of various quiz datasets.

Function List

• Data Enhancement Module: Convert documents into data-enhanced datasets.
• Parameter Training Module: Train additional LoRA parameters to generate a parameterized representation of the document.
• inference module: Merge parameterized representations of related documents and insert them into the LLM for inference.
• Environment Installation: Provides detailed steps and dependencies for installing the environment.
• self-improvement: Supports direct use of pre-enhanced data files or self-processed data enhancements.
• Search preparation: Download and prepare Wikipedia datasets for retrieval.

Using Help

Environment Installation

1. Create and activate a virtual environment:

   conda create -n prag python=3.10.4
conda activate prag

2. Install the necessary dependencies:

   pip install torch==2.1.0
pip install -r requirements.txt

3. modifications src/root_dir_path.py hit the nail on the head ROOT_DIR variable is the address of the folder where the PRAG is stored.

data enhancement

1. Use pre-enhanced data files:

   tar -xzvf data_aug.tar.gz

2. Self-processing data enhancement:
   • Download the Wikipedia dataset: bash
mkdir -p data/dpr
wget -O data/dpr/psgs_w100.tsv.gz https://dl.fbaipublicfiles.com/dpr/wikipedia_split/psgs_w100.tsv.gz

   • intend BM25 Retrieve: bash
# Please refer to the project documentation for specific steps.


parametric training

1. Generate a parameterized representation of the document:

   # Please refer to the project documentation for specific steps.

inference

1. Parameterized representations of related documents are merged and inserted into the LLM for inference:

   # Please refer to the project documentation for specific steps.