zotvault 0.9.2

Local-first paper pipeline orchestrator between Zotero and Obsidian: detect, note, fetch PDF, queue AI analysis, index.

ZotVault

Local-first paper pipeline orchestrator between Zotero and Obsidian.

ZotVault watches your Zotero library and, for every new paper, automatically: creates an Obsidian note, secures a PDF (open-access first, politely rate-limited), puts the paper on an AI-analysis queue, and keeps your vault's index current. Around that core loop it adds one-shot DOI/arXiv adding, paper search, a local dashboard, arXiv keyword alerts, an in-library citation graph, embedding-based related-paper suggestions, and synthesis-cluster proposals.

Everything runs on your machine. No cloud accounts required, no API keys required for the core loop, zero runtime dependencies (Python ≥ 3.9 standard library only).

[search]                       [collect]                [archive]           [wiki]
arXiv / S2 / Crossref  ──▶  dashboard / CLI / agent ──▶  Zotero (BBT)  ──▶  Obsidian vault
        ▲                        │ you approve             │ sqlite poll        │
        └── arXiv alert inbox ───┘                         ▼                    ▼
                                            daemon: note → PDF (OA→proxy) → queue → index
                                                     + citation graph · related · synthesis

Features

• One-shot add — zotvault add 10.1103/PhysRevB.1.1 arXiv:2405.01234: metadata via Crossref/DataCite/arXiv → straight into Zotero through the same local channel the browser connector uses — including the OA PDF when one exists (Zotero downloads it itself, like the browser connector). Duplicates are detected before saving; a pipeline cycle runs immediately after the add. Optional translation-server support for arbitrary URL imports.
• Automatic wiki-fication — new Zotero items become Obsidian notes (template-compatible, atomic writes). Existing notes are never rewritten; your manual sections are structurally safe.
• Edit-safe highlight sync — Zotero PDF highlights land in ONE marker-delimited block per note (grouped by color, deep links back to the exact annotation), kept in sync incl. deletions. Figure/area annotations are embedded as images (copied from Zotero's cache). Color groups can be renamed to your semantics (label_red = "Core Claims"). Everything outside the block is untouchable; unmarked legacy notes are opt-in ([annotations] adopt_existing). See docs/MIGRATION.md.
• PDF resolution, politely — Zotero attachment → cache → arXiv → Unpaywall → (opt-in) institutional proxy with browser-session cookies. Sequential, delayed, daily-capped: designed to not get your campus blocked. See docs/PROXY.md.
• AI-analysis queue — ZotVault doesn't run an LLM; it feeds yours. zotvault queue --json lists unanalyzed papers with readable PDF paths; when your agent (e.g. Claude batch) writes *_analysis.md, completion is auto-detected.
• Dashboard — zotvault web → http://127.0.0.1:8377 : search → tick → add, queue, alerts inbox, suggestions, audit trail. Localhost only.
• arXiv alerts — daily keyword digest into a review inbox. Nothing enters Zotero without your click (propose, don't execute).
• Alert triage (assist) — optionally, a small local Ollama model scores inbox candidates 0–10 for relevance under a strict JSON contract (constrained decoding + validation + one retry); the dashboard sorts by score. Advisory only, off by default.
• Citation graph — Semantic Scholar citation counts + who-cites-whom within your library, regenerated into Citation_Graph.md.
• Related papers — local Ollama embeddings over your analysis notes → Related_Suggestions.md link candidates. Free, offline.
• Synthesis suggestions — clusters analyzed-but-unsynthesized papers into proposed review topics.
• Auditable — every automatic action lands in a SQLite trace (zotvault trace).

Requirements

• Python ≥ 3.9 (CLI is cross-platform; the double-click app + install-daemon autostart are macOS today)
• Zotero desktop with Better BibTeX — required: it is ZotVault's citekey source. Without it, items can't be named and nothing syncs (ZotVault will mark them blocked and tell you).
• An Obsidian vault (any folder). ZotVault writes per-paper notes under <papers_subdir>/<citekey>/ — configure the paths; it does not impose a vault structure.
• Optional: pdftotext (poppler) for full-text AI analysis, Ollama (related papers / local analysis), an institutional web proxy for licensed PDFs.

Quick start

pip install zotvault                   # or: pipx install zotvault / uv tool install zotvault
# (from source: git clone https://github.com/jbaek-dev/ZotVault && pip install ./ZotVault)
zotvault init                          # interactive setup: paths, email, vault scaffolding + doctor
# (--yes for the non-interactive template; add [alerts] keywords later if you want digests)
zotvault run-once --dry-run
zotvault run-once

Run it continuously. Two options on every platform:

• System tray (recommended): pip install "zotvault[tray]" then zotvault tray — daemon + tray icon with Open Dashboard / Run now / Pause / Quit. The only third-party packages ZotVault can use, and only if you opt in.
• OS service: zotvault install-daemon — macOS writes a launchd plist, Linux writes a systemd user unit, Windows prints the Task Scheduler command. Never auto-loads; it tells you the enable command.

macOS extra: double-click ZotVault.app (bash scripts/build_app.sh).

Every command also works without installing, via python3 -m zotvault.cli <cmd>.

Commands

command	what it does
init / doctor	config file / environment health check
run-once [--dry-run] / daemon	one cycle / poll loop (+dashboard thread)
install-daemon	autostart: launchd plist (macOS) / systemd unit (Linux) / schtasks cmd (Windows)
tray	daemon + system-tray icon (needs pip install "zotvault[tray]")
add <ids…> [--dry-run]	resolve DOI/arXiv/URL → save to Zotero
search <query> [--source arxiv|s2|crossref]	search with in-library marks
web	dashboard server in the foreground
queue [--json]	papers awaiting AI analysis
alerts [--fetch|--approve N|--dismiss N]	arXiv digest inbox
enrich [--limit N]	citation graph + embeddings + suggestion notes
assist [--limit N]	score pending alerts with the small local model
related <citekey> / synthesis [--write]	similarity / cluster proposals
status / trace [--limit N]	state summary / audit trail

Design guarantees

• Read-only toward Zotero's data. Library reads use a temp snapshot of zotero.sqlite; adds go through Zotero's own connector endpoint (Zotero writes its DB itself). ZotVault never touches storage/; downloaded PDFs live in ~/.zotvault/pdfs/.
• Vault safety. Existing notes are never rewritten; auto-notes (Citation_Graph.md, Related_Suggestions.md, _Synthesis_Suggestions.md) are clearly marked and ZotVault-owned; index.md is only patched via a strict counter regex; log.md is append-only; --dry-run previews; no delete code path exists.
• Polite networking. OA-first, sequential fetches, delays, hard daily caps (separate, stricter cap for the proxy), honest User-Agent, backs off on HTTP 429.
• Nothing personal in code. All user-specific values live in ~/.zotvault/config.toml.

Agent integration

Any agent can drive ZotVault through the localhost API (/api/search, /api/add, /api/queue, /api/status). A ready-made tool module for the Polaris local agent (not yet public) lives in that repo (polaris/tools/zotvault_tools.py): search results show up with in-library marks, and adding requires the user to have explicitly picked papers.

Tests

python3 -m unittest discover -s tests -v   # 101 tests, no network needed

License

MIT — see LICENSE. Changelog in CHANGELOG.md.