research-hub-pipeline 0.33.1

CLI + MCP server for Zotero + Obsidian + NotebookLM research pipelines. Run `research-hub init` after install.

research-hub

Zotero + Obsidian + NotebookLM, wired together for AI agents.

繁體中文說明 → README.zh-TW.md

---

What this is

A CLI + MCP server that does three things at once:

1. Ingest academic papers into Zotero (citations) + Obsidian (structured notes) + NotebookLM (briefings) — one command.
2. Organize papers into clusters, sub-topics, and an Obsidian graph coloured by research label.
3. Serve 52 MCP tools so Claude Code / Codex / any MCP-compatible AI can drive the whole thing.

Built for PhD students and research teams who already use AI agents daily and don't want to context-switch between six tabs.

Source code vs vault

research-hub has two separate locations on your computer. This is intentional:

	Source code	Vault
What	The Python package + CLI	Your research data
Where	site-packages/research_hub/ (managed by pip)	~/knowledge-base/ (default, you choose during init)
Contains	CLI, MCP server, dashboard renderer	Paper notes, Obsidian graph, crystals, Zotero sync
Shared?	Yes — same package for every user	No — each user has their own vault

After pip install, run research-hub init to create your vault. If you already have an Obsidian vault, point init at it — research-hub adds its folders alongside your existing notes without overwriting anything.

Run research-hub where at any time to see exactly where your config and vault live.

What makes it different

1. Crystals — pre-computed answers, not lazy retrieval (v0.28)

Every RAG system, including Karpathy's "LLM wiki", still assembles context at query time. research-hub's answer: store the AI's reasoning, not the inputs.

For each research cluster, you generate ~10 canonical Q→A "crystals" once (via emit/apply, using any LLM you like). When an AI agent asks "what's the SOTA in X?", it reads a pre-written 100-word paragraph — not 20 paper abstracts.

research-hub crystal emit --cluster llm-agents-software-engineering > prompt.md
# Feed prompt.md to Claude/GPT/Gemini, save answer as crystals.json
research-hub crystal apply --cluster llm-agents-software-engineering --scored crystals.json

Token cost per cluster-level query: ~1 KB (crystal read) vs ~30 KB (cluster digest). 30× compression without losing quality, because the quality was pre-computed.

→ Why this is not RAG

2. Live dashboard with direct execution (v0.27)

research-hub serve --dashboard

Opens a localhost HTTP dashboard at http://127.0.0.1:8765/. Every Manage-tab button directly executes the CLI command instead of copying to clipboard. Vault changes push to the browser via Server-Sent Events. Fallback to static clipboard mode when the server isn't running.

3. Obsidian graph auto-coloured by label (v0.27)

research-hub vault graph-colors --refresh

Writes 14 colour groups to .obsidian/graph.json: 5 per cluster path + 9 per paper label (seed, core, method, benchmark, survey, application, tangential, deprecated, archived). Every research-hub dashboard run auto-refreshes them. Open Obsidian Graph View — your vault is visually structured by meaning, not just file tree.

4. Sub-topic-aware Library + citation-graph cluster split (v0.27)

Big clusters (331 papers?) don't render as a flat list anymore. They're grouped by sub-topic, each expandable. And if your cluster has no sub-topics yet:

research-hub clusters analyze --cluster my-big-cluster --split-suggestion

Uses Semantic Scholar citation graph + networkx community detection to suggest 3-8 coherent sub-topics. Writes a markdown report you review before running topic apply-assignments.

---

Install

pip install research-hub-pipeline
research-hub init              # interactive config + vault layout
research-hub serve --dashboard # opens browser

Python 3.10+. No OpenAI/Anthropic API key required — research-hub is provider-agnostic (all AI generation uses emit/apply pattern; you feed prompts to your own AI).

For Claude Code / Claude Desktop users

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "research-hub": {
      "command": "research-hub",
      "args": ["serve"]
    }
  }
}

Then talk to Claude:

"Claude, add arxiv 2310.06770 to a new cluster called LLM-SE" "Claude, generate crystals for the LLM-SE cluster" "Claude, what's this cluster about?" → Claude calls list_crystals + read_crystal → gets the pre-written 100-word answer

52 MCP tools cover: paper ingest, cluster CRUD, labels, quotes, draft composition, citation graph, NotebookLM, crystal generation, fit-check, autofill.

Quickstart (5 commands)

# 1. Initialize vault
research-hub init

# 2. Ingest one paper
research-hub add 10.48550/arxiv.2310.06770 --cluster llm-agents

# 3. Open live dashboard
research-hub serve --dashboard

# 4. Generate crystals once you have a few papers
research-hub crystal emit --cluster llm-agents > prompt.md
# (feed prompt.md to your AI, save response as crystals.json)
research-hub crystal apply --cluster llm-agents --scored crystals.json

# 5. Ask your AI questions — it reads crystals, not papers
# (via Claude Desktop MCP, or any MCP-compatible client)

Status

• Latest: v0.33.1 (2026-04-17)
• Tests: 1249 passing, 14 skipped, 2 xfail baselines (documented search-quality issues)
• Platforms: Windows, macOS, Linux
• Python: 3.10+
• Dependencies: pyzotero, pyyaml, requests, rapidfuzz, networkx, platformdirs (all pure-Python)
• Optional: playwright extra for NotebookLM browser automation

Architecture docs

• MCP tools reference — all 50+ tools categorized + signatures
• Example Claude Desktop flow — worked example: ingest → crystallize → query
• Import folder — local file ingest for analyst persona (PDF/DOCX/MD/TXT/URL)
• Anti-RAG crystals — why pre-computed Q→A beats retrieval
• Upgrade guide — migrating from older versions
• Task-level workflows — v0.33+ 5 MCP wrappers (ask/brief/sync/compose/collect) that collapse 3-4 call sequences into 1
• Screenshot workflow — re-render any dashboard tab via dashboard --screenshot CLI
• Audit reports — audit_v0.26.md … audit_v0.33.md
• NotebookLM setup — CDP attach flow + troubleshooting
• Papers input schema — ingestion pipeline reference

Workflow reference

Stage	Command	What it does
Init	init / doctor	First-time config + health check
Find	search / verify / discover new	Multi-backend paper search + DOI resolution + AI-scored discovery
Ingest	add / ingest	One-shot or bulk paper ingest into Zotero + Obsidian
Organize	clusters new/list/show/bind/merge/split/rename/delete	Cluster CRUD
Topic	topic scaffold/propose/assign/build	Sub-topic notes from subtopics: frontmatter
Label	label / find --label / paper prune	Canonical label vocabulary (seed/core/method/...)
Crystal	crystal emit/apply/list/read/check	Pre-computed canonical Q→A answers
Analyze	clusters analyze --split-suggestion	Citation-graph community detection for big clusters
Sync	sync status / pipeline repair	Detect + repair Zotero ↔ Obsidian drift
Dashboard	dashboard / serve --dashboard / vault graph-colors	Static HTML or live HTTP server + auto-refresh Obsidian graph
NotebookLM	notebooklm bundle/upload/generate/download	Browser-automated NLM flows (CDP attach)
Write	quote / compose-draft / cite	Quote capture, markdown draft assembly, BibTeX export

Two personas

Persona	Install	Zotero?	Best for
Researcher (default)	pip install research-hub-pipeline[playwright]	Yes	PhD students, academic literature review
Analyst	research-hub init --persona analyst	No — Obsidian only	Industry research, white papers, technical docs

Both personas get the same dashboard, MCP server, and crystal system.

For developers

git clone https://github.com/WenyuChiou/research-hub.git
cd research-hub
pip install -e '.[dev,playwright]'
python -m pytest -q  # 1249 passing

Package name on PyPI: research-hub-pipeline CLI entry point: research-hub

License

MIT. See LICENSE.