Local Deep Research — Privacy-First AI Research Agent

Introduction

Local Deep Research is an open-source AI research tool that conducts multi-step investigations using LLMs and multiple search engines. It can run entirely on local hardware with Ollama or llama.cpp, or connect to cloud providers like Google or Anthropic. The project focuses on achieving high accuracy while keeping all data processing private and encrypted.

What Local Deep Research Does

• Conducts iterative multi-step research by breaking questions into sub-queries
• Searches across multiple engines including arXiv, PubMed, Brave, and SearXNG
• Indexes and searches private documents stored locally for RAG-based answers
• Generates structured research reports with citations and source references
• Supports both local LLMs via Ollama and cloud LLM providers

Architecture Overview

The system operates as a Python application with a Flask-based web interface. When a research query is submitted, it decomposes the question into sub-queries, dispatches them to configured search backends, retrieves and ranks results, then synthesizes findings through the configured LLM. Local document search uses vector embeddings stored in a local database. All processing can remain on the user machine when paired with a local LLM.

Self-Hosting & Configuration

• Install via pip and launch the web interface with a single command
• Configure LLM provider settings for Ollama, llama.cpp, or cloud endpoints
• Enable or disable search engines (arXiv, PubMed, Brave, SearXNG) as needed
• Point the local document search to directories containing your private files
• Optionally encrypt local data stores for additional privacy protection

Key Features

• Multi-engine search combining academic, web, and private document sources
• Full local execution possible with no data leaving the machine
• Iterative research strategy that refines queries based on intermediate findings
• Structured report generation with inline citations
• Configurable LLM backend supporting both local and cloud models

Comparison with Similar Tools

• Perplexica — web search focused; Local Deep Research adds academic databases and private documents
• Khoj — general AI assistant; Local Deep Research specializes in deep iterative research
• GPT Researcher — cloud-dependent; Local Deep Research supports fully local execution
• RAGFlow — document QA focused; Local Deep Research adds multi-engine web and academic search
• SearXNG — metasearch only without LLM synthesis; Local Deep Research adds AI-driven analysis

FAQ

Q: What LLMs work best with Local Deep Research? A: Models with strong reasoning capabilities work best. The project documentation lists recommended models for different hardware configurations.

Q: Can I use it without internet access? A: Yes, when using a local LLM and searching only local documents. Web and academic searches require internet.

Q: How does it handle large document collections? A: Documents are chunked and embedded into a local vector store for efficient semantic retrieval.

Q: What search engines are supported? A: arXiv, PubMed, Brave Search, SearXNG, and configurable additional sources.

Sources

• https://github.com/LearningCircuit/local-deep-research