AI Research Assistant com RAG Multimodal
MohdFaisal2165/AI-Research-Assistant
RAG local com comparação 4-way de LLMs e dicionado por voz
Ver no GitHubSistema full-stack de Retrieval-Augmented Generation (RAG) que ingere PDFs acadêmicos localmente e os processa com Llama 3.2, comparando resultados em tempo real com Claude, GPT-4o e Gemini. Inclui interface React com dicionado por voz via Whisper e feedback automático para benchmark de modelos.
Instalação
npx convez-ai MohdFaisal2165/AI-Research-Assistant1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise comparativa de respostas entre modelos LLM em documentos acadêmicos
- Pesquisa e extração de informações de PDFs densos com interface conversacional
- Benchmark de performance de modelos open-source vs cloud-based em contexto real
Veredito

Ferramenta sofisticada para pesquisadores que precisam comparar modelos LLM em contexto RAG. Exige setup complexo (Ollama, APIs, Node.js) mas oferece insights valiosos sobre hallucinations e performance relativa. Bem executada para casos acadêmicos e benchmarking rigoroso.
README
A tradução para o português deste README ainda está na fila.
🔬 AI Research Assistant
A full-stack, multimodal Retrieval-Augmented Generation (RAG) system built to parse dense academic PDFs and converse interactively using local and cloud-based Large Language Models.
This project enables real-time model comparison by routing user queries simultaneously across four powerful LLMs: Llama 3.2 (Local), Claude 3.5 Sonnet, GPT-4o, and Gemini Flash.
✨ Features
- Local RAG Pipeline: Fully automated ingestion pipeline that reads academic PDFs, extracts semantic chunks using
SentenceTransformers, and natively clusters them in a localizedChromaDB. - 4-Way Model Comparison: Automatically query a single prompt to Claude, GPT-4o, Gemini, and Local Llama simultaneously to benchmark their understanding of the context.
- Strict Hallucination Prevention: Specialized prompts designed to strictly bottleneck local open-source models directly to the PDF sources provided, preventing hallucinated citations.
- Audio Dictation: Seamless Voice-to-Text inference powered by Whisper AI.
- Automated Feedback Loop: Rate model outputs natively within the React UI and log responses locally to an Excel Database for fine-tuning insights.
🛠 Tech Stack
- Backend: Python, FastAPI, Uvicorn, LangChain, Ollama
- Frontend: React, Vite, CSS Native Glassmorphism
- Vector Database: ChromaDB
- Embedding Agent:
all-MiniLM-L6-v2(sentence-transformers)
🚀 Setup & Installation
1. Prerequisites
- Python 3.10+
- Node.js (v18+)
- Local instance of Ollama running
llama3.2model.
ollama run llama3.2
2. Environment Variables
Create a file named .env in the root directory (you can use the provided .env.example as a template) and add your API keys:
OPENAI_API_KEY="sk-..."
ANTHROPIC_API_KEY="sk-ant-..."
GEMINI_API_KEY="AIza..."
3. Backend Setup
Create your virtual environment and install dependencies:
python -m venv .venv
source .venv/bin/activate # On Windows: .venv\Scripts\activate
pip install -r requirements.txt
4. Automated Data Ingestion (Vector DB)
To utilize the Retrieval-Augmented Generation limits, you must construct the vector database.
- Place any research PDFs you want to chat with into the
data/raw_papers/directory. - Run the ingestion pipeline to automatically chunk and embed the papers into ChromaDB:
python backend/ingest_pdfs.py
Note: This script will inherently skip indexing pages like standard Tables of Contents.
💻 Running the Application
You need two terminal windows actively running.
Terminal 1 (FastAPI Backend):
# Ensure your virtual environment is active!
python backend/api.py
Terminal 2 (React Frontend):
cd frontend
npm install
npm run dev
The application will be accessible via http://localhost:5173. Ask complex questions and instantly benchmark outputs across four of the world's most powerful LLMs!