Local Deep Research is an AI-powered research assistant designed to perform deep, agentic research using multiple language models and search engines while maintaining full local control and encryption. The project is written in Python and addresses a critical gap in open-source research tools by achieving approximately 95% accuracy on the SimpleQA benchmark using only a single RTX 3090 GPU running Qwen3.6-27B, making it the first open-source project to report such performance on fully local hardware. The repository also reports 77% accuracy on xbench-DeepSearch benchmarks.
The core functionality allows users to ask complex research questions, and the system automatically conducts research across multiple sources including the web, academic papers from arXiv and PubMed, Semantic Scholar, and user-owned private documents. The tool synthesizes findings into reports with proper citations. Users can choose between quick pipeline modes for fast factual lookups or fully agentic deep research strategies for complex analysis. The LangGraph Agent Strategy represents the most advanced mode, where the language model autonomously decides which specialized search engines to use, adapts its search strategy based on findings, and collects significantly more sources than pipeline-based approaches.
A distinctive feature is the knowledge base building capability, where research sessions automatically identify valuable sources that users can download directly into their encrypted library. The system extracts text from academic papers, articles, and web pages, indexes everything, and makes it searchable. This allows users to compound their knowledge over time by asking questions across both their personal document collection and live web sources simultaneously.
Security and privacy form the foundation of the project's design. All user data remains encrypted using AES-256 encryption via SQLCipher, with encryption keys derived from user passwords rather than stored separately. Each user receives an isolated encrypted database, and per-user API keys are stored encrypted within personal databases rather than in shared server-level storage. The Docker setup implements strict security practices including capability dropping, no-new-privileges restrictions, and non-root runtime execution. Docker images are signed with Cosign using GitHub's keyless OIDC flow and include SLSA provenance attestations and SPDX SBOMs. The project contains no telemetry, analytics, or tracking of any kind.
The repository supports flexible deployment options including Docker, Docker Compose, pip installation, and specialized deployments like Unraid. It works with all local and cloud language models including llama.cpp, Ollama, Google, OpenAI, Mistral, and Anthropic endpoints. The project maintains comprehensive security scanning through OpenSSF Scorecard, CodeQL, Semgrep, DevSkim, Bearer, OSV-Scanner, npm-audit, Retire.js, container security tools, and OWASP ZAP.
The project overlaps with contributors from github/gh-aw, solo-io/gloo, and longhorn/longhorn repositories.
