Asia/Kolkata
About
Work
ProjectsJanuary 15, 2025

Deep Researcher RAG System

Vijay Makkad
Deep Researcher RAG System

Overview

Deep Researcher is a multi-agent research pipeline that turns a single natural-language query into a comprehensive, cited research report. Instead of relying on one LLM call, the system orchestrates specialized agents for web retrieval, document chunking, vector search, and final synthesis—producing answers grounded in retrieved evidence rather than model hallucination. The project demonstrates how RAG (Retrieval-Augmented Generation) can scale beyond simple Q&A into a full research workflow. Users submit a topic through a Streamlit interface and receive structured output with sources, making it useful for literature reviews, competitive analysis, and technical deep dives.

Tech Stack

The backend is built in Python with LangChain for agent orchestration and prompt chaining. Retrieved documents are embedded and indexed in a FAISS vector store for fast similarity search. Recursive web search agents gather fresh content from the open web, while an LLM synthesis layer combines findings into a coherent narrative. The frontend is a Streamlit app that keeps the interaction simple while the multi-agent pipeline runs behind the scenes.

Highlights

  • Designed a multi-agent architecture where retrieval, ranking, and synthesis are handled by separate specialized steps.
  • Integrated FAISS for efficient vector similarity search over large document collections.
  • Implemented recursive web search to expand coverage beyond a single query result page.
  • Built a Streamlit UI that accepts a research question and returns a formatted report with source references.

More projects

TensorFlow

Video Action Recognition

CNN-based action recognition model using MovieNet architecture for Samsung Prism. Achieved 97% training accuracy on compressed video datasets.
Case study
React.js

MingleMate

Real-time chat app supporting 100+ concurrent users with Tailwind CSS, DaisyUI, secure JWT auth, and optimized MongoDB queries.
Case studyView project