Local Deep Research

Local Deep Research (LDR) is a self-hostable research assistant that takes a complex question, searches across the web, academic sources, and your own documents, and synthesizes the findings into a report with citations. It runs against local models through Ollama or any OpenAI-compatible endpoint as well as cloud providers, and pairs with a SearXNG instance for web search. Research modes range from a quick summary (30 seconds to 3 minutes) through detailed research and full report generation to document analysis over private collections.

The langgraph-agent strategy is an autonomous mode in which the model decides what to search, which specialized engines to use (arXiv, PubMed, Semantic Scholar, and others), and when to synthesize, switching engines based on what it finds. Community benchmarks using this strategy with Serper search and local Ollama models report the following results, with the caveats of small samples, LLM-grader noise, and possible SimpleQA contamination in newer base models.

Model	SimpleQA	xbench-DeepSearch
Qwen3.6-27B	95.7% (287/300)	77.0% (77/100)
Qwen3.5-9B	91.2% (182/200)	59.0% (59/100)
gpt-oss-20B	85.4% (295/346)	–

Sources found during research can be downloaded into an encrypted per-user library, where LDR extracts text, indexes it, and makes it searchable alongside live web results in later sessions. Each user has an isolated SQLCipher database encrypted with AES-256 and keyed from the user's password, which is never stored; per-user LLM API keys live inside that database. The application makes no telemetry or analytics calls; the only network traffic is the searches, LLM calls, and optional Apprise notifications you configure. It is available as Docker images (signed with Cosign, with SLSA provenance and SBOMs), Docker Compose, or a pip package, and is reachable through a web UI, a Python API, an authenticated HTTP API, a CLI, and an MCP server for Claude Desktop and Claude Code.

Features

• Research modes: quick summary, detailed research, report generation with sections and table of contents, and document analysis
• Agentic strategy: langgraph-agent lets the model choose engines and decide when to synthesize
• Search sources: arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, Elasticsearch, Wayback Machine, The Guardian, and Wikinews, plus Tavily, Google, and Brave as paid options
• Knowledge base: download sources into an encrypted library that is indexed and searchable in future research
• LangChain retrievers: any vector store (FAISS, Chroma, Pinecone, Weaviate, Elasticsearch) can act as a search engine
• Per-user encryption: isolated SQLCipher databases with password-derived keys and encrypted API keys
• Journal quality system: reputation scoring over 212K+ indexed sources with predatory-journal detection, using OpenAlex, DOAJ, and Stop Predatory Journals data
• Chat mode and streaming: multi-turn research conversations with WebSocket progress updates
• Subscriptions: scheduled research digests on topics or queries, delivered daily, weekly, or on a custom schedule
• MCP server: search, quickresearch, detailedresearch, generatereport, analyzedocuments, and configuration tools for local STDIO use
• Analytics and benchmarking: a cost and usage dashboard and a CLI for running SimpleQA-style benchmarks
• Export and history: PDF or Markdown export and saved, searchable research history