Local Deep Research — Précis

What it is

Local Deep Research (LDR) is an open-source, MIT-licensed Python web app for "deep, agentic research": you ask a complex question, and it searches the web, academic databases, and your own documents, then synthesizes a cited report. Its defining promise is privacy and control — it runs entirely locally (Ollama + SearXNG), stores every user's data in a per-user AES-256-encrypted SQLCipher database, and makes no telemetry calls. It ships as Docker images, pip install local-deep-research, or a Flask web UI on localhost:5000.

The headline result: the first open-source project to report ~95% SimpleQA (n=500) and 77% xbench-DeepSearch on a single RTX 3090 running Qwen3.6-27B fully local, using its langgraph-agent strategy.

How it works

• Research engine — AdvancedSearchSystem (search_system.py) coordinates a pluggable strategy (BaseSearchStrategy). Five user-facing strategies: source-based, focused-iteration (×2), topic-organization, and langgraph-agent (autonomous tool-calling).
• langgraph-agent — the flagship: LangChain's create_agent() builds a tool-calling agent whose tools are search engines themselves. It spawns parallel subagents via a research_subtopic tool, each with its own search/fetch tools, and a thread-safe SearchResultsCollector dedupes sources across subagents.
• Search engines — a plugin system over BaseSearchEngine, ~25 engines (arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, Tavily, Brave, GitHub, Wayback…), plus local collections and LangChain retrievers.
• Knowledge base — research findings feed an encrypted "library": download sources, extract text, chunk + embed (sentence-transformers + FAISS), then search collections as engines next time ("your knowledge compounds over time").
• Security — a DLP-style egress guardrail (ADR-0007) classifies every source and sink on two axes (Sensitivity × Exposure) and forbids sensitive data from reaching exposing sinks. Per-user encrypted databases, CSRF, SSRF/metadata-IP blocking, and 22+ CI security scanners.

~189K lines of Python. Python ≥3.12, requires AVX-capable CPUs for the bundled scientific wheels.

Distinctive angles

• Egress as a first-class architecture concern — an in-process PDP (security/egress/policy.py) with PEPs at every boundary, a PEP-578 audit hook that classifies raw socket.connect targets, and honest documentation that it is "a correctness guardrail, NOT a hard security boundary."
• Per-user encrypted DB threading — one SQLAlchemy QueuePool per user (not one global pool), because each user has their own SQLCipher file with a unique key; the FD-budget analysis and cleanup layers are documented in docs/architecture.md.
• Benchmark-driven — a community-maintained HuggingFace leaderboard tracks accuracy across models, engines, and strategies, so users can compare local models before downloading weights.

Source: raw/local-deep-research — verbatim README. Analysis from full repo clone.