local-deep-research: Research with Local or Cloud AI

Overview

Local Deep Research is a research assistant that searches multiple sources, then uses an LLM to synthesize findings into answers and reports with citations. It is designed to reduce the manual work of gathering and cross-checking material across the web, academic indexes, and a personal document library.

You can run it with local models and self-hosted services, or connect cloud providers. That flexibility makes it useful both for people prioritizing privacy and for users who prefer hosted models or search services. Its encrypted per-user database and configurable egress controls offer safeguards, but do not guarantee that every request remains private.

Key Features

• Agentic research strategy that can choose search engines as it gathers sources, alongside faster research modes.
• Search across web, academic services such as arXiv and PubMed, and configured private documents.
• Generate summaries and structured reports with citations, and export results as Markdown or PDF.
• Connect local models through Ollama, LM Studio, or llama.cpp, as well as cloud and compatible API providers.
• Build a searchable document library from downloaded sources and use it alongside live search.
• Provide a web interface, Python and HTTP APIs, and an MCP server for local assistant integrations.
• Store user data in separate SQLCipher-encrypted databases and configure research egress scopes.

Use Cases

• Researchers can find and synthesize papers across academic search sources, then keep useful material in a searchable library.
• Analysts and journalists can produce sourced background reports and revisit previous research.
• Privacy-conscious individuals can run local models and keep research history on a self-hosted installation, while checking which services their searches use.
• Teams with internal documentation can combine a configured document retriever with external research when their data and egress policies allow it.

Project Facts

• Language: Python
• License: MIT
• Stars: 9.1k
• Forks: 833
• Topics: academia, anthropic, arxiv, brave, deep-research, encryption, home-automation, homeserver, local, local-deep-research, local-llm, mistral, ollama, openai, pubmed, research, research-tool, retrieval-augmented-generation, searxng, self-hosted
• Archived: No

Getting Started

For a quick pip installation, run:

pip install local-deep-research
python -m local_deep_research.web.app

You will also need a configured LLM endpoint and search service. See the README for Docker, Compose, and setup instructions.

Considerations

• Running fully locally requires suitable local model-serving software and hardware; the README describes a single RTX 3090 setup for its reported benchmark, but hardware needs vary by model.
• Local inference does not make web searches private. Search engines and model endpoints may receive queries, and a self-hosted SearXNG can forward them to upstream services.
• Egress controls are described as best-effort guard rails, not an air gap or hard security boundary. Strong isolation may require container- or firewall-level restrictions.
• The project supports many providers and deployment paths, so initial configuration can involve choosing models, search engines, and privacy settings. Consult the security and installation documentation before exposing services beyond local use.