This project aims to reimplement the paper “RAG: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”. For model training, I explored three approaches:
- Fine-tuning both the retriever and generator using pre-trained models.
- Training both the retriever and generator from scratch.
- Training only the retriever component.
Due to limitations in GPU resources and time constraints, it proved impractical to train a fully functional model from scratch or to fine-tune both the retriever and generator concurrently. Consequently, I opted to utilize LangChain, a Python library tailored for constructing RAG systems, to develop the RAG application.
The resulting application, Note Helper, is a web-based tool that allows users to select a folder of documents as the knowledge base. Users can add more documents to the knowledge base at any time and ask questions about the content.
Note Helper is deployed locally, utilizing Ollama's open-source LLM models or OpenAI's models for versatile performance. It realized a context management system, which allows the model to have long-term memory. It features a user-friendly web application for seamless interaction and is complemented by a PyQt5 desktop application, providing users with flexible options for managing and querying their document-based knowledge base.
For a more comprehensive understanding of the project, please refer to the Note Helper README located in the note-helper-rag directory and the Training README found in the train directory.