~ 95% 在 SimpleQA 上 (例如 Qwen3.6-27B 在 3090 上)。支持所有本地和云端的 LLM (LLaMA.cpp、Ollama、Google 等)。10 多个搜索引擎 - arXiv、PubMed、您的私人搜索引擎
~ 95% 在 SimpleQA 上 (例如 Qwen3.6-27B 在 3090 上)。支持所有本地和云端的 LLM (LLaMA.cpp、Ollama、Google 等)。10 多个搜索引擎 - arXiv、PubMed、您的私人搜索引擎
AI research assistant you control. Run locally for privacy, use any LLM and build your own searchable knowledge base. You own your data and see exactly how it works.
Option 1: Docker Run (Linux)
…
Mac / Windows / WSL2 users:
--network hostonly works on native Linux. On Docker Desktop it silently fails to publish port 5000 and leaveslocalhostpointing at the LDR container itself (so it can't reach Ollama/SearXNG). Use Option 2 below, or see the Windows/WSL2 FAQ entry for a workingdocker runrecipe.
Option 2: Docker Compose
CPU-only (all platforms):
curl -O https://raw.githubusercontent.com/LearningCircuit/local-deep-research/main/docker-compose.yml && docker compose up -d
With NVIDIA GPU (Linux):
curl -O https://raw.githubusercontent.com/LearningCircuit/local-deep-research/main/docker-compose.yml && \
curl -O https://raw.githubusercontent.com/LearningCircuit/local-deep-research/main/docker-compose.gpu.override.yml && \
docker compose -f docker-compose.yml -f docker-compose.gpu.override.yml up -d
Open http://localhost:5000 after ~30 seconds. For GPU setup, environment variables, and more, see the Docker Compose Guide.
Option 3: pip install
pip install local-deep-research
python -m local_deep_research.web.app # starts the web UI on http://localhost:5000
You'll also need Ollama (or any OpenAI-compatible LLM endpoint) and SearXNG running — see the pip install guide for the full recipe. Works on Windows, macOS, and Linux. SQLCipher encryption is included via pre-built wheels — no compilation needed. PDF export on Windows requires Pango (setup guide). If you encounter issues with encryption, set
export LDR_BOOTSTRAP_ALLOW_UNENCRYPTED=trueto use standard SQLite instead.
Detailed install guides: Docker · Docker Compose · pip · Unraid · full install reference
Older CPU (x86-64)? LDR needs an AVX-capable CPU — Intel Sandy Bridge / AMD Bulldozer (2011) or newer. Several scientific Python dependencies (pandas, scikit-learn) ship wheels that crash with
Illegal instructionon older CPUs. ARM64 (aarch64) is fully supported. Every release is smoke-tested against this floor, including AVX-without-AVX2 CPUs (#4480).
You ask a complex question. Local Deep Research (LDR):
Choose the research strategy that fits: quick pipeline modes for fast facts, or fully agentic deep research for complex analysis and academic work.
LangGraph Agent Strategy — An autonomous agentic research mode where the LLM decides what to search, which specialized engines to use (arXiv, PubMed, Semantic Scholar, etc.), and when to synthesize. It adaptively switches between search engines based on what it finds and collects significantly more sources than pipeline-based strategies — this is the strategy behind the ~95% SimpleQA result above. Select langgraph-agent in Settings.
flowchart LR
R[Research] --> D[Download Sources]
D --> L[(Library)]
L --> I[Index & Embed]
I --> S[Search Your Docs]
S -.-> R
Every research session finds valuable sources. Download them directly into your encrypted library — academic papers from ArXiv, PubMed articles, web pages. LDR extracts text, indexes everything, and makes it searchable. Next time you research, ask questions across your own documents and the live web together. Your knowledge compounds over time.
flowchart LR
U1[User A] --> D1[(Encrypted DB)]
U2[User B] --> D2[(Encrypted DB)]
Your data stays yours. Each user gets their own isolated SQLCipher database encrypted with AES-256, with the key derived from your password. Your password is never stored — login works by attempting to decrypt your database, so the database files on their own are unusable to anyone who obtains them. Per-user LLM API keys live encrypted inside the same personal database rather than in a shared server-level store.
The Docker setup ships with cap_drop: ALL, no-new-privileges, and a non-root runtime, with the bundled Ollama and SearXNG images pinned by digest. Or run fully local with Ollama + SearXNG and nothing ever leaves your machine.
In-memory credentials: Like any application that uses secrets at runtime, credentials are held in process memory during active sessions — mitigated with session-scoped credential lifetimes and core dump exclusion. See the Security Policy for the full threat model.
Supply Chain Security: Docker images are signed with Cosign using GitHub's keyless OIDC flow, include SLSA provenance attestations, and ship with attested SPDX SBOMs. See Verifying images and SBOMs for the step-by-step verification commands.
Security Transparency: Scanner suppressions are documented with justifications in Security Alerts Assessment, Scorecard Compliance, Container CVE Suppressions, and SAST Rule Rationale. Some alerts (Dependabot, code scanning) can only be dismissed or are very difficult to suppress outside the GitHub Security tab, so the files above do not cover every dismissed finding.
Detailed Architecture → | Security Policy → | Security Review Process →
Local Deep Research contains no telemetry, no analytics, and no tracking. We do not collect, transmit, or store any data about you or your usage. No analytics SDKs, no phone-home calls, no crash reporting, no external scripts. Usage metrics stay in your local encrypted database.
The only network calls LDR makes are ones you initiate: search queries (to engines you configure), LLM API calls (to your chosen provider), and notifications (only if you set up Apprise).
Since we don't collect any usage data, we rely on you to tell us what works, what's broken, and what you'd like to see next — bug reports, feature ideas, and even which features you love or never use all help us improve LDR.
Headline results from the community benchmarks using the langgraph-agent strategy with Serper search, fully local via Ollama:
| 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) | – |
Caveats: small samples, LLM-grader noise, and SimpleQA contamination risk on newer base models.
Picking a local model? The same community-maintained dataset tracks accuracy across models, search engines, and research strategies — the fastest way to see which Ollama / LM Studio / llama.cpp models actually work well for deep research before you download multi-GB weights. Browse the full leaderboard on Hugging Face →
Submit your own results → (contributors are listed in CONTRIBUTORS.md), or run benchmarks locally →.
LDR respects robots.txt and identifies itself honestly when fetching web pages — no stealth or anti-detection techniques. In rare cases this means a page that blocks automated access won't be fetched, which we consider the right trade-off.
from local_deep_research.api import LDRClient, quick_query
# Option 1: Simplest - one line research
summary = quick_query("username", "password", "What is quantum computing?")
print(summary)
# Option 2: Client for multiple operations
client = LDRClient()
client.login("username", "password")
result = client.quick_research("What are the latest advances in quantum computing?")
print(result["summary"])
The code example below shows the basic API structure - for working examples, see the link below
…
Ready-to-use HTTP API Examples → examples/api_usage/http/
暂无开放 Issues，或尚未同步最近议题。