DefinePK

Research Objective

To develop a web application that leverages Retrieval Augmented Generation (RAG) to enable users to converse with a Large Language Model (LLM) in the context of their uploaded documents, thereby improving document comprehension and information retrieval.

Methodology

The research proposes a web application with several components: Authentication (Auth), Chat, User Management, and Feedback. The system converts user documents into text chunks, generates embeddings, and stores them in a vector database. User queries are used to retrieve relevant text chunks, which are then augmented with the user's prompt and fed to an LLM (Llama 3.2 1B) for contextual responses. The development environment utilizes Typescript, Langchain.js, NodeJS, and Ollama, with Visual Studio Code as the IDE and Git for version control. The application is deployed on Google Cloud Platform (GCP) using Ubuntu, Nginx, and Express.js for the backend, and Netlify for the frontend. Testing involves unit, integration, and end-to-end tests for both frontend and backend.

Methodology Flowchart

                        graph TD;
    A[User Uploads Document] --> B[Document Processing: Chunking & Embedding];
    B --> C[Store in Vector DB];
    D[User Query] --> E[Retrieve Relevant Context from Vector DB];
    E --> F[Augment Prompt with Context & Query];
    F --> G[LLM Generates Response];
    G --> H[Display Response to User];
    I[User Feedback] --> J[Feedback Storage];
    K[Admin Actions] --> L[User Management];                    

Discussion

The paper discusses potential improvements to the RAG workflow, including advanced chunking methods, metadata integration, handling structured data, query rewriting and decomposition, query-to-document expansion, routing subsystems, re-ranking mechanisms, context consolidation, and multi-modality. Practical implications highlight improved user interaction with documents, efficient information retrieval, and enhanced user experience through summarization and generative tasks.

Key Findings

The developed web application successfully enables users to interact with LLMs in the context of their own documents, overcoming limitations of LLMs regarding recent information and private data. The RAG approach enhances accuracy by grounding responses in provided context and broadens LLM applicability by allowing interaction with private data.

Conclusion

The proposed LLM-based web application with RAG effectively addresses the limitations of traditional LLMs by providing contextual understanding of user-provided documents. Future work includes handling picture-rich PDFs, adding a download option for questions, and conducting tests with diverse user groups.

Fact Check

* The paper was published in the International Journal of Innovations in Science & Technology (IJIST), Vol. 06 Issue. 04 in December 2024. (Confirmed by citation and publication date)
* The application uses the Llama 3.2 1B LLM. (Confirmed in Methodology section)
* The development environment uses Typescript, Langchain.js (version 0.3.7), NodeJS (version 22.12.0), and Ollama (version 0.3.4). (Confirmed in Development Environment section)

Mind Map