Llama4_DeepSeek_RAG: A Multi-Model Comparison PDF Intelligent Q&A System

Llama4_DeepSeek_RAG is a PDF intelligent Q&A application based on Retrieval-Augmented Generation (RAG) technology. Its core feature is supporting parallel comparison of dual models Llama-4 and DeepSeek-R1. Users can upload PDF documents for natural language Q&A, and intuitively compare the reasoning processes and answer quality of different models, which is suitable for model selection and RAG effect evaluation. The project is maintained by skhaneefa42, open-sourced on GitHub, and uses Streamlit to build an interactive interface.

• Original author/maintainer: skhaneefa42
• Source platform: GitHub
• Original title: Llama4_DeepSeek_RAG
• Original link: https://github.com/skhaneefa42/Llama4_DeepSeek_RAG
• Release date: 2026-05-30

This project aims to address the need of developers and researchers for multi-model performance comparison, providing an intuitive RAG application evaluation tool.

Core Features

1. Dual Model Support: Integrates Llama-4 (general-purpose multilingual, instruction-following) and DeepSeek-R1 (inference-specialized, chain-of-thought output). Users can flexibly select or compare them in parallel.
2. Intelligent PDF Parsing: The process is document parsing → text chunking → vector embedding → semantic retrieval, preserving document structure and achieving precise matching.
3. Streamlit Interface: Supports drag-and-drop PDF upload, dialogue interaction, model switching, and result display.

Technical Architecture

RAG pipeline: PDF upload → text extraction → chunk processing → vector embedding → vector storage; User query → query vectorization → semantic retrieval → context assembly → model inference → answer generation. Semantic retrieval is based on vector embedding technology, which can understand synonyms and perform cross-language retrieval (depending on the capability of the embedding model).

The system is designed with a multi-dimensional comparison mechanism to help users evaluate model performance:

1. Answer Accuracy: Compare the matching degree between the model's answer and the document content;
2. Reasoning Transparency: Chain-of-thought display of DeepSeek-R1 vs direct answer of Llama-4;
3. Response Speed: Differences in inference efficiency between different models;
4. Answer Style: Formality, detail level, structure level, etc.

Users can call both models simultaneously to intuitively observe the differences across dimensions.

Typical scenarios of this application include:

1. Enterprise Document Q&A: Import internal materials such as product manuals and technical documents to build an enterprise knowledge assistant;
2. Academic Research Assistance: Upload papers to quickly extract key information, verify cited content, and improve literature research efficiency;
3. Model Selection Evaluation: Compare the performance of the two models on real business data to assist deployment decisions;
4. Education and Training: Show the differences in thinking styles of different models to help students understand AI technology.

Llama4_DeepSeek_RAG is not only a practical RAG tool but also an open-source model comparison research platform. It lowers the technical threshold for multi-model evaluation, allowing individual developers and small and medium-sized enterprises to conduct professional model capability evaluations. With the development of the open-source large model ecosystem, such comparison tools will help the community better utilize model advantages and promote the implementation and optimization of AI applications.