Local Research OS: Autonomous RAG Agent

A privacy-first, air-gapped AI system that autonomously monitors ArXiv for new research, ingests papers into a local Vector Database, and provides a citation-aware chat interface. Built to solve the "Context Loss" and "Source Hallucination" problems common in LLM retrieval.

Architecture

• The Scout (Autonomous Agent): Runs daily, queries ArXiv API, uses Llama 3 to score abstracts (1-10) for relevance, and auto-downloads high-value papers.
• The Ingestion Engine (ETL Pipeline): A batch-processing workflow in n8n that splits PDFs, injects "Source Stamps" into every text chunk (to guarantee citation accuracy), and embeds them into Qdrant.
• The Brain (RAG Chat): A local Llama 3.1 instance connected via Open WebUI that performs semantic search with a strict "Cite Sources" system prompt.

Tech Stack

• Orchestration: n8n (Dockerized)
• Vector Database: Qdrant (Local)
• Inference Engine: Ollama (running Llama 3.1 8B)
• Interface: Open WebUI
• Scripting: JavaScript (Custom Node logic for metadata injection)

Getting Started

Prerequisites

• Docker Desktop (4GB+ RAM allocated)
• Ollama installed locally

Installation

1. Clone the repo:

   git clone [https://github.com/VS251/local-research-os.git](https://github.com/VS251/local-research-os.git)

2. Start the stack:

   docker-compose up -d

3. Import the workflows from the /workflows folder into n8n (localhost:5678).

Key Features

• Zero Data Ingress: 100% local. No data is sent to OpenAI or Anthropic.
• Source Stamping: Custom ETL logic injects SOURCE_FILENAME: {name} into every vector embedding, solving the "Lost Header" problem in RAG retrieval.
• Automatic Curation: The system filters noise by "reading" abstracts for you before downloading.