research-hub-pipeline 0.43.0

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

research-hub

Build your research cluster once. Ask AI about it thousands of times. Zotero + Obsidian + NotebookLM, wired together for AI agents.

繁體中文說明 → README.zh-TW.md

---

The 10-minute story

Say you want to get up to speed on "harness engineering for LLM agents" — a subfield that barely existed 6 months ago. Traditional workflow: search arXiv, skim abstracts, manually note key claims, fight Obsidian, occasionally wish you had a RAG. 2 hours.

research-hub workflow:

research-hub clusters new --query "LLM evaluation harness" --slug llm-evaluation-harness
research-hub search "language model evaluation harness" --to-papers-input \
    --cluster llm-evaluation-harness > papers.json
research-hub ingest --cluster llm-evaluation-harness --no-verify

3 minutes later your vault has 6 key papers with structured notes. Now push them to NotebookLM and pull back an auto-generated briefing:

research-hub notebooklm bundle   --cluster llm-evaluation-harness
research-hub notebooklm upload   --cluster llm-evaluation-harness   # Chrome CDP-attach, no API key
research-hub notebooklm generate --cluster llm-evaluation-harness --type brief
research-hub notebooklm download --cluster llm-evaluation-harness --type brief

Your vault now has .research_hub/artifacts/llm-evaluation-harness/brief-*.txt — a ~300-character synthesis covering all 6 papers, generated by NotebookLM from the uploaded sources (no prompt engineering needed, no headless-browser hacks — it attaches to your existing Chrome session).

2 more minutes: generate the AI summary layer (crystals) and the structured entity/claim registry (memory):

research-hub crystal emit --cluster llm-evaluation-harness > prompt.md
# (paste prompt to Claude/GPT, save response as crystals.json)
research-hub crystal apply --cluster llm-evaluation-harness --scored crystals.json

research-hub memory emit  --cluster llm-evaluation-harness > mem-prompt.md
research-hub memory apply --cluster llm-evaluation-harness --scored memory.json

Now open Claude Desktop and ask:

You: "What's the current SOTA in LLM evaluation harness?" Claude (via MCP): calls read_crystal("llm-evaluation-harness", "sota-and-open-problems") → gets a pre-written 180-word answer with paper citations. ~1 KB read, 0 abstracts fetched at query time.

That pre-written answer is a crystal. You paid the reasoning cost once; every subsequent question is ~1 KB of cached analysis. See hub/llm-evaluation-harness/crystals/ in your vault for the 10 canonical Q&As generated above.

---

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 cluster you generate ~10 canonical Q→A crystals once, using any LLM you like. When an AI agent later asks "what's the SOTA in X?", it reads a pre-written paragraph — not 20 paper abstracts. Token cost per query: ~1 KB (crystal read) vs ~30 KB (cluster digest). 30× compression.

Because the quality was pre-computed, it doesn't degrade at query time. See the harness-engineering example crystals — one folder, 10 Q&As answering "what is this field?", "what are the main threads?", "where do experts disagree?", "what's SOTA?", etc.

→ Why this is not RAG

2. Structured memory layer — entities, claims, methods (v0.36)

Crystals store prose. Memory stores the underlying structure: named entities (benchmarks, models, concepts), typed claims with confidence + supporting papers, and method taxonomies. For the harness cluster:

hub/llm-evaluation-harness/memory.json
├── 14 entities  (vla-eval, SafeHarness, M*, LIBERO, SEC-bench, ...)
├── 12 claims    ("Harness is locus of progress", "Specialized beats generic +22%", ...)
└── 7 methods    (reflective code evolution, lifecycle-integrated defense, ...)

AI agents query entities via list_entities, claims via list_claims(min_confidence="high"), methods via list_methods. No RAG over prose — structured lookup over structured data.

3. 4 personas, 1 codebase, dashboard adapts (v0.38)

Same vault, 4 rendered dashboards:

Persona	Install	Dashboard vocabulary	Hidden tabs
Researcher (PhD STEM, Zotero)	pip install research-hub-pipeline[playwright,secrets]	Cluster / Crystal / Paper / Citation graph	(none)
Humanities (Zotero, quote-heavy)	pip install research-hub-pipeline[playwright,secrets]	Theme / Synthesis / Source	(none)
Analyst (industry, no Zotero)	pip install research-hub-pipeline[import,secrets]	Topic / AI Brief / Document	Diagnostics, Bind-Zotero
Internal KM (lab / company)	pip install research-hub-pipeline[import,secrets]	Project area / AI Brief / Document	Diagnostics, Bind-Zotero

Side-by-side screenshots: docs/personas.md. Your first 10 minutes guide →

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

research-hub serve --dashboard

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

5. Cluster integrity + 100% orphan coverage (v0.37 + v0.39)

Papers drift, rebind v2 catches it. On the maintainer's 1063-orphan vault: 33% → 100% coverage via 8-heuristic chain + auto-create-from-folder proposals.

research-hub doctor                       # catches 12+ classes of drift
research-hub clusters rebind --emit       # proposes 80%+ assignments
research-hub clusters rebind --apply report.md --auto-create-new

→ 6 failure modes × 4 personas mitigation matrix

---

Install

# Researcher / Humanities (use Zotero + NotebookLM)
pip install research-hub-pipeline[playwright,secrets]

# Analyst / Internal KM (no Zotero, import local files)
pip install research-hub-pipeline[import,secrets]

research-hub init              # 4-option interactive persona prompt
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"]
    }
  }
}

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

Then talk to Claude:

"Claude, what's in my llm-evaluation-harness cluster?" → read_crystal("what-is-this-field") → 180-word answer "Claude, which claims have high confidence?" → list_claims(cluster="llm-evaluation-harness", min_confidence="high") → 10 structured claims with paper refs "Claude, add arxiv 2310.06770 to LLM-SE cluster" → add_paper(...) → Zotero + Obsidian + NotebookLM entries

---

Status

• Latest: v0.41.0 (2026-04-19)
• Tests: 1423 passing, 15 skipped, 3 xfailed (CI: Linux + Windows + macOS × Python 3.10/3.11/3.12)
• 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

• Your first 10 minutes — guided tour for each of the 4 personas
• User personas — 4 persona profiles with per-persona feature matrix
• Cluster integrity — 6 failure modes + mitigation matrix across all 4 personas
• MCP tools reference — all 60 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)
• Screenshot workflow — re-render any dashboard tab
• Audit reports — audit_v0.26.md … audit_v0.41.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 (doctor has 12+ checks, --autofix for mechanical backfills)
Find	search / verify / discover new	Multi-backend paper search + DOI resolution + AI-scored discovery
Ingest	add / ingest / import-folder	One-shot or bulk paper ingest into Zotero + Obsidian
Organize	clusters new/list/show/bind/merge/split/rename/delete/rebind/scaffold-missing	Cluster CRUD + 8-heuristic rebind + hub scaffolding
Topic	topic scaffold/propose/assign/build	Sub-topic notes from subtopics: frontmatter
Label	label / find --label / paper prune / paper lookup-doi	Canonical label vocabulary + Crossref DOI backfill
Crystal	crystal emit/apply/list/read/check	Pre-computed canonical Q→A answers
Memory	memory emit/apply/list/read	Structured entities/claims/methods registry
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

For developers

git clone https://github.com/WenyuChiou/research-hub.git
cd research-hub
pip install -e '.[dev,playwright]'
python -m pytest -q  # 1423 passing

Contributing: see CONTRIBUTING.md. Reporting security issues: see SECURITY.md.

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

License

MIT. See LICENSE.