Multilingual RAG-Powered Research Assistant with Paper Recommendations

14825 Generative AI and LLM Final Project

Finding relevant academic research can be time-consuming, especially across domains and languages. This project aimed to build a multilingual research assistant using a Retrieval-Augmented Generation (RAG) pipeline to synthesize domain-specific knowledge and recommend academic papers. The goal was to streamline early-stage research by combining semantic search with LLM-generated summaries in the user’s preferred language.

I built a web application using Streamlit that allows users to enter natural language queries, select a research domain, and optionally receive paper recommendations. The application supports 5–10 languages, ensuring that output matches the input language. I used FlowiseAI and LangChain to design the RAG pipeline and orchestrate the components of the system. I embedded 20 academic papers into ChromaDB for domain-specific semantic search. When a user submits a query, the system retrieves relevant passages and combines them with LLM-generated responses to produce a synthesized answer. I also compared multiple embedding models and large language models to evaluate pipeline performance. The application was deployed to the cloud and made publicly accessible.

The assistant successfully handled multilingual inputs and produced high-quality summaries and relevant paper links across multiple domains. Comparing two embedding models and two LLMs revealed that model selection significantly impacted both relevance and coherence of responses. The pipeline demonstrated strong performance in retrieving and generating responses that matched user intent, while the cloud deployment ensured accessibility and responsiveness.

User Interface for project

This project showcased how RAG systems can be adapted for practical, multilingual academic use. It emphasized the importance of embedding model quality, LLM selection, and orchestration tools like LangChain in building flexible, modular pipelines. Using LangChain enabled clean chaining of retrieval, generation, and response formatting steps. Building the app strengthened my skills in LLM integration, vector search, and web deployment. Future improvements could include scaling the paper corpus, supporting real-time paper ingestion, and integrating translation capabilities for cross-lingual search.