DIY deep research that actually runs on your GPU

What it does

Local Deep Research is a self-hosted research assistant that takes a complex question, searches across web and academic sources (arXiv, PubMed, Semantic Scholar, etc.), and synthesizes a cited report. It supports both local LLMs via Ollama/llama.cpp and cloud providers, with 20+ research strategies ranging from quick facts to deep agentic exploration.

The interesting bit

The project explicitly targets the “can I actually run this at home?” crowd. It claims to be the first open-source tool hitting ~95% on SimpleQA (n=500) with a single RTX 3090 running Qwen3.6-27B. The langgraph-agent strategy is the engine behind that number: an autonomous agent that decides which search engines to query, when to keep digging, and when to synthesize—collecting more sources than fixed pipelines.

Key highlights

• Compounding knowledge base: Download sources from any research session into an encrypted personal library; future queries search your accumulated documents plus the live web
• Per-user AES-256 encryption via SQLCipher with no password recovery—admins can’t read your data
• Extensive security posture: Cosign-signed Docker images, plus automated scanning via CodeQL, Semgrep, OWASP ZAP, OSV-Scanner, and others
• Broad LLM support: Ollama, llama.cpp, OpenAI-compatible APIs, Google, Anthropic, etc.
• 10+ search engines including SearXNG, Brave, arXiv, PubMed; also indexes your private documents

Caveats

• Docker --network host fails silently on macOS/Windows Docker Desktop, breaking port publishing and localhost resolution; the README warns you need workarounds
• The ~95% SimpleQA claim is specifically for the langgraph-agent strategy with Qwen3.6-27B; your mileage will vary with other models and strategies
• In-memory credentials remain plaintext during active sessions, which the documentation candidly notes is an industry-wide limitation, not a solved problem

Verdict

Worth a look if you want research automation without shipping your queries to OpenAI or Perplexity. Skip it if you need turnkey SaaS reliability or don’t want to babysit Ollama and SearXNG.