LearningCircuit/local-deep-research

Local Deep Research – What It Is

Local Deep Research (LDR) is an open‑source, self‑hosted AI research assistant. It lets you ask complex questions and automatically gathers information from the web, academic sources, and your own documents, then synthesizes a citation‑rich report. All processing runs on your own hardware (e.g., a single RTX 3090) and data stays encrypted on your machine, so you retain full privacy and control.

---

Core Capabilities

Capability	How It Works	What You Get
Agentic research	Uses a LangGraph‑based autonomous agent that decides which search engines to query (arXiv, PubMed, Semantic Scholar, etc.) and when to synthesize results.	Deep, multi‑step analysis that can achieve ~95 % accuracy on the SimpleQA benchmark.
Search‑plus‑library	Downloads sources into an encrypted SQLite (SQLCipher) library, extracts text, creates embeddings, and indexes them for fast local search.	A growing personal knowledge base that can be queried alongside live web searches.
Multiple LLM back‑ends	Works with any OpenAI‑compatible endpoint (Ollama, OpenAI, Azure, etc.) – you provide the model you want, such as Qwen‑3.6‑27B.	Flexibility to run fully local models or use cloud APIs while still keeping data private.
Egress control	Built‑in “egress scopes” (Adaptive, Public only, Private only, Strict, Unprotected) restrict where network traffic may go, with best‑effort socket guards and SSRF checks.	Fine‑grained privacy settings that limit accidental data leaks.
Encrypted per‑user storage	Each user gets an AES‑256‑encrypted SQLCipher database; API keys are stored inside the same encrypted file.	Credentials and research history are never written in clear text.
Web UI & API	A Flask‑based web interface (http://localhost:5000) plus REST endpoints for programmatic use.	Interactive research sessions or integration into other tools.

---

Getting Started (Three Easy Paths)

1. Docker (Linux) – Pull the Ollama model server, a SearXNG search engine, then the LDR container. The example in the README wires them together with --network host.
2. Docker‑Compose – One‑command setup for CPU‑only or GPU‑accelerated deployments (the latter uses the NVIDIA runtime). Includes a ready‑made docker‑compose.yml and a GPU override file.
3. pip install – pip install local-deep-research and run python -m local_deep_research.web.app. You still need Ollama (or another LLM endpoint) and SearXNG running, but this works on Windows/macOS as well.

All three methods come with detailed docs (Docker, Docker‑Compose, pip, Unraid) that cover environment variables, hardware requirements (AVX‑capable CPU, RTX 3090 recommended), and troubleshooting.

---

Security & Privacy Highlights

• Zero telemetry – No analytics, crash reporters, or phone‑home calls.
• Encrypted storage – SQLCipher (AES‑256) protects each user’s database; the decryption key is derived from the password and never persisted.
• Supply‑chain hardening – Docker images are signed with Cosign, include SLSA provenance, and ship SPDX SBOMs. The repo runs extensive CI security scans (CodeQL, Semgrep, DevSkim, OWASP ZAP, etc.).
• Runtime hardening – Containers drop all capabilities, run non‑root, and enforce no-new-privileges. In‑memory credentials are cleared on session end and core‑dump protection is enabled.
• Egress policy – Configurable scopes limit outbound traffic; Strict and Private only modes add socket‑level guards against accidental public requests.

---

Who Might Use It?

• Researchers & students who need a reproducible, private way to gather literature and web sources.
• Developers building custom AI agents who want a ready‑made pipeline for search → retrieval → synthesis.
• Privacy‑conscious power users who want the capabilities of ChatGPT‑style agents without sending their queries to third‑party servers.

---

Quick Verdict

Local Deep Research is a fully‑featured, locally‑run AI research assistant that combines multi‑LLM orchestration, secure document ingestion, and strong privacy controls. It’s a solid choice if you have a capable GPU (or at least an AVX‑compatible CPU) and want to keep both data and inference on‑premise.