OVP2 — grounded knowledge for your Obsidian vault

Turn the articles you save into a knowledge base you can ask — and every answer cites a real line in a real source.

OVP2 is a local-first app for Obsidian vaults. It reads what you capture, extracts grounded memory, crystallizes durable claims, and gives you a portal to browse, search, and ask — without inventing citations.

简体中文 · Install details · Operator runbook

Knowledge graph: themes and claims from a real dogfood vault.

Most note tools store text. OVP2 keeps a truth layer:

If a statement cannot point at evidence in your vault, it does not become durable knowledge. Search, the graph, and Ask are projections of that ledger — rebuildable anytime.

ovp2 serve (or the OVP2 desktop app) opens a localhost portal over your vault.

A morning dashboard: captures, reads, new claims, and items that need attention.

Browse by collection and month. Open any source for memory cards, the full original document, and claims that cite it — with a neighborhood graph on the side. Markdown renders properly, Mermaid diagrams included.

Per source, on demand: a deep summary, a 中文 translation, companion links to related sources, and chat about this one — enrichment you ask for, not a bill the daily loop runs up on everything you ever saved.

Durable vs caveated claims, grouped by theme. Switch List / Graph / Terrain when you want structure instead of a spreadsheet of notes.

Each theme also gets a topic page: its durable claims woven into prose, where every sentence cites a claim key. A draft that cannot cite is repaired once, then rejected — the page never ships ungrounded.

Ask in natural language. The agent searches claims, sources, and evidence cards; the Process panel shows what it touched; the answer carries numbered citations you can open.

Sources, claims, packs, themes — and the full text of every source body, not just titles. ⌘K / Ctrl+K from anywhere.

The interface has always had a 中文 locale. Beyond that, OVP2 can carry a Chinese projection of the knowledge itself: claims, memory cards, and topic pages get zh versions alongside the English ones.

It is a projection, deliberately — the ledger stays English and single-authority, and the zh layer is rebuildable from it, so a translation can never become a second source of truth that quietly disagrees. New material is queued as it arrives and old material backfills in the background, both inside the vault's own token budget.

Also in the portal: Tags, Entities, Work queue (what enrichment is running, its pace and ETA), System (runs, doctor, LLM settings, schedule). Light Atelier and dark Vault themes.

The daily loop: capture sweep → grounded read per new source → ledgers → rebuild the read model. Crystal synthesis turns reader packs into cross-source claims behind mechanical gates.

Some things you save are entry points, not articles — a brand homepage, a gallery, a docs index. Tag the capture ovp_skip and the pipeline drops it at intake, before it fetches or spends anything. Remove the tag later and it is picked up again on the next sweep; nothing is deleted. ovp_force is the other direction: read this one even though it is too short to clear the automatic gate.

We tried to detect these pages automatically and measured that it does not work — on a real 1,448-source corpus the structural signals flagged a 73k-character engineering writeup and a 46k-character CUDA article next to the three genuine navigation pages. Whether a bookmark is worth reading is your call, so it stays a tag.

Prebuilt CLI for macOS arm64 and Linux x64 (current line: v2.0.1). No Rust toolchain required.

curl --proto '=https' --tlsv1.2 -LsSf \
  https://github.com/fakechris/obsidian_vault_pipeline/releases/latest/download/ovp-cli-installer.sh | sh

or:

brew install fakechris/ovp2/ovp2

ovp2 --version

Full channels, desktop DMG, and rollback notes: docs/install.md.

1. LLM credentials (for live reads / Ask) — write <vault>/.ovp/providers.toml, which the app reads itself. Shell env still wins when set, so one-off overrides keep working.

   [env]
   ANTHROPIC_API_KEY = "sk-ant-..."
   OVP_LLM_TIMEOUT_SECS = "480"
   
   [budget]
   daily_token_budget = 2000000   # optional; reported by `ovp2 usage`, not enforced

   The portal's System → LLM settings page edits this file for you (keys masked once saved). A scheduled job cannot reliably shell-source an env file — that is why this is a config file and not daily.env.

2. One daily pass (try --dry-run first):

   ovp2 daily --vault-root ~/Documents/my-vault --client live

3. Schedule (optional):

   ovp2 schedule install --vault-root ~/Documents/my-vault

4. Portal:

   ovp2 serve --vault-root ~/Documents/my-vault

5. Pinboard (optional): PINBOARD_TOKEN=user:TOKEN then
   ovp2 pinboard-sync --vault-root … --live --max 200

Local-first: product state is plain files under your vault (.ovp/ ledgers + notes). No account, no telemetry.

Only leave the machine when you configure them:

• LLM calls — text you process is sent to the provider behind your API key (or local endpoint). No key → offline/replay only.
• Pinboard — only with --live and your token (never logged).
• Web / GitHub enrichment — fetches bookmarked URLs (and repo metadata for GitHub links) when enabled. Set XQUIK_API_KEY to resolve X/Twitter status bookmarks through Xquik.
• Manual diagnostic compare — only if you run the compare command against an external service you choose; not part of daily.
• Publishing — ovp2 publish is the only command that pushes anything outward, only to a repo you name, and only durable claims.

What it costs is visible too: every metered LLM call lands in .ovp/usage/, and ovp2 usage reports tokens by day and by lane against your budget line. The budget is soft — it reports, it does not cut you off mid-run. Per-run limits are the throttle (--max-sources, the enrichment queue's own cap).

Xquik is an independent third-party service. Not affiliated with X Corp. "Twitter" and "X" are trademarks of X Corp.

Screenshots in docs/images/ were taken from a local dogfood vault (public tech clippings). Re-capture anytime with the portal running; review for secrets before publishing.

Rust workspace (CLI + portal + optional desktop). The daily loop, crystal synthesis, topic pages, Ask agent, source enrichment, the 中文 projection, and the portal all run on a real dogfood vault daily — currently ~1,450 sources and ~1,500 durable + caveated claims. See releases for the latest artifacts.

Dual-licensed under either of

• MIT license (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)

at your option. Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Exception: the vendored IBM Plex web fonts (console-ui/src/design/fonts/) remain under the SIL Open Font License 1.1 — see console-ui/src/design/fonts/LICENSE.txt.