_{Crisper, as MP4  ·  every number in it is measured, see ARCHITECTURE.md}

Smart Related Notes

A left-sidebar panel that surfaces the notes most semantically similar to the one you're reading, so you browse your vault by meaning, not by folders. Open any note and the panel ranks the rest by how closely they relate, as a stack of cards you can click to jump to.

It runs on a small multilingual embedding model that lives entirely on your machine: no cloud, no API key, no second app or server. After a one-time model download it works fully offline, and it understands German, English and 100+ other languages, so matches cross languages naturally.

Contents  ·  What's new  ·  Roadmap  ·  What it does  ·  How it works  ·  Settings  ·  Install

What's new in 3.0

Ranking stopped being about wording alone. The panel now also reads the shape of your own links, which is what produced every number above. Full detail lives in ARCHITECTURE.md; the short version is that similarity and your link graph disagree often enough to be worth fusing, and roughly a third of the links you make point somewhere similarity structurally cannot look.

Roadmap

Next, in 4.0. Asking the vault questions, and letting other tools use what it knows:

• Search by category: ask for "characters" and get the characters, not the note about characters, with no tags and no map-of-content required. Measured at 0.78 member-precision, but only on a model that can handle short queries; on a plain paraphrase model the same pipeline tops out at 0.42, so the feature will only offer itself where it works rather than quietly returning noise.
• Semantic views in Bases: meaning as a column you can sort, filter and group on.
• Link suggestions with receipts: not "these look similar" but the exact detail two notes share, so you can judge it at a glance. The underlying gate stack was measured on a related study (judged precision 0.40 to 0.60 at n=15, 95% interval 0.36 to 0.80); the feature itself has not been measured yet.
• A ranker built from channels: each signal scores independently, with the weights fitted to your own links, per vault, on your machine. The admission rule is the interesting part: a new channel only earns a place if it is measurably independent of the ones already there. That is how the link-graph channel was chosen. Its rank correlation with the content channel is 0.075 and it added 0.08 recall; a keyword channel correlated at 0.30 and made things worse.

Exploring, for 5.0. Hypotheses with open questions rather than promises:

• The signal only your machine has: which notes you open and edit in one sitting is evidence neither the text nor the links can give, and it is the last plausibly independent channel left. It is also the one a cloud service structurally cannot have, because that data never leaves your computer.
• Vaults of a hundred thousand notes: every ranking currently scans the whole vault. A real nearest-neighbour index turns that into a lookup. Unglamorous, and the thing that decides whether any of the above survives a serious archive.
• Notes that disagree, and help while you write.

Ideas that fail their measurement get cut rather than shipped quietly. Bridges, analogy search, a duplicate alarm, a personal similarity metric and a keyword channel were each built, measured, and dropped.

What it does

Finding what relates

• Semantic ranking: for the active note, ranks every other note by meaning and shows the top matches as cards: title, muted folder path, a short snippet, and a similarity % pill. Click a card to open that note. With no note open, the panel lists your recent notes instead of sitting empty.
• Link-graph ranking: the panel does not rank by wording alone. When two notes share context you have already connected them through, that counts as evidence they belong together, and such a card says via the note that bridges them. This is what lifted held-out link recall from 0.66 to 0.75.
• Semantic search: the magnifier in the panel header ranks your whole vault by meaning against a typed query, not just keyword matches.
• Linked-notes mode: the link icon switches the cards to show what the current note links to, the structural complement to similarity ranking.

Building your graph

• Inline link suggestions: when you mention a concept that already has a note, it glows with a slim underline; one click turns the mention into a wikilink. It is context-aware, so a common word like "analysis" only glows where it fits the topic, and it works with or without the easy-links plugin.
• Smarter [[ completion: typing [[ ranks existing notes by how well they fit what you are writing, and offers to create a new note for a strongly relevant concept that does not have one yet.
• Vault insights: a whole-vault report of suggested links, orphans, near duplicates, stale notes and suggested tags. See below.

Seeing your vault

• Vault map: every note as a point, placed so notes about similar things sit together, coloured by cluster and labelled automatically from the notes themselves. Click a point to open it, click a cluster to hide it. Open it with the Open the vault map command.
• Surprising connections: pairs your links connect through shared context but whose wording is too different for similarity to ever pair them. Because each note must sit outside the other's top matches to qualify, this list cannot fill up with obvious siblings the way a similarity ranking does.

Private, and it stays that way

• Fully local: embeddings run in-app on the CPU (WASM, multi-threaded; WebGPU optional). Notes are never sent anywhere, and it works offline after the one-time download.
• Multilingual: matches notes across German, English and 100+ other languages.
• Persisted index: vectors are saved to disk, so reopening the vault is instant.
• Incremental updates: changed, created and renamed notes are re-embedded on a 20-second idle pause, so typing never kicks off work mid-edit.
• Keyword fallback: while the index is still building, the panel falls back to a cheap keyword ranking (shown with a ~ pill) so it is never empty.
• Clear status: a live status line shows progress, and surfaces an error rather than silently showing nothing.

Vault insights

Run "Vault insights (suggested links, orphans, duplicates)" from the command palette to generate a report note for the whole vault:

• Suggested links: the strongest related notes that you have not linked yet, ranked by similarity. The fastest way to grow a sparse graph.
• Surprising connections: pairs your links connect through shared context but whose wording is too different for similarity to ever pair them. Each note sits outside the other's top matches, so this list cannot fill up with obvious siblings the way a similarity ranking does. Every entry names the note that bridges the two.
• Suggested tags: notes that are missing a tag most of their semantic neighbours share. The plugin infers a likely category (e.g. a character profile that lacks your goa/character tag) from similarity alone, only proposing discriminative tags.
• Orphan notes: notes with no links in or out, each paired with its closest relative as a starting point.
• Possibly duplicate: near-identical pairs worth merging or cross-linking.
• Stale notes: the oldest-edited notes, for review.

It writes/refreshes Vault Insights (Smart Related Notes).md and opens it; every entry is a clickable wikilink. Nothing is changed in your notes.

How it works

Each Markdown note is split into short passages (windows that fit the model's context), covering the whole note, and every passage is turned into a vector by the embedding model. Adjacent passages are then grouped into coherent ideas (~200-500 words, at heading and topic boundaries, with short atomic notes kept whole), so a note is represented at three levels: an overall vector, its idea vectors, and its passage vectors. For the note you're viewing, every other note is ranked by the best cosine similarity of their passages (with its title weighted), blended with how strongly the two notes share a whole idea, so a note related by one coherent idea spanning several paragraphs surfaces, not just one lucky passage match. The closest matches are shown as cards with a similarity percentage.

The model runs through @huggingface/transformers on the local ONNX runtime, on the CPU via WASM by default, multi-threaded so a full reindex is quick, and memory-stable. (A WebGPU option exists and is faster, but its GPU backend can accumulate memory on large vaults, so it's an explicit opt-in.) The whole engine lives in its own worker, which is shut down after an idle period (configurable) — so when you're just reading and browsing, the model's hundreds of MB to several GB are returned to the system instead of sitting resident; it reloads transparently in a few seconds when next needed. An Indexing speed setting trades CPU threads against reindex time. The only network traffic ever is one-time: the model weights from the Hugging Face Hub, and the ONNX runtime .wasm (version-pinned), which is cached into the plugin's own folder; afterwards nothing is fetched again and the plugin works offline.

Vectors persist as compact JSON in the plugin's config dir, so the index survives restarts and only changed notes are re-embedded.

For the full picture, see ARCHITECTURE.md: the multi-granularity embeddings, the multi-stage ranking funnel, the measured model A/B, mean-centering, the structural channel that ranks by your link graph, template de-crowding, and where tag-free concept search stands.

Settings

Quality and speed

• Performance profile: one-click presets. Balanced is lighter and faster; Best quality uses a larger model and more context for the strongest matches.
• Model: the embedding model (a dropdown of vetted choices, or paste a custom Hugging Face id). The default paraphrase-multilingual-MiniLM-L12-v2 is a symmetric sentence-similarity model, the right tool for ranking how alike two notes are. paraphrase-multilingual-mpnet-base-v2 (Best quality) is stronger but larger. e5 models are retrieval-oriented and rank note similarity less well.
• Compute device: Auto (recommended) runs on the CPU (WASM), which is memory-stable. WebGPU is faster but its GPU backend can accumulate memory on large vaults, so it's an explicit opt-in. Switching re-downloads the model.
• Indexing speed: how many CPU threads embed notes. Fast (all cores, quickest), Balanced (default), Light (1 thread, slowest). A CPU/speed knob — the engine's memory is dominated by the loaded model itself, not the thread count. Editing a note stays fast at any setting.
• Unload model when idle: after this long without indexing or searching (default 15 minutes; "Right after use" unloads 30 seconds after the last embed), the embedding engine is shut down and its memory returned to the system. It reloads automatically in a few seconds on the next use — no re-download. Ranking when switching notes never needs the engine and is unaffected.

The panel

• Number of results: how many cards to show.
• Minimum similarity: hide matches below this topical-similarity score (0-1). Scores are mean-centered (the embedding noise floor is removed), so unrelated notes sit near 0 and ~0.2 cleanly separates on-topic notes. Raise for a tighter list.
• Show snippet: toggle the per-card text preview.

Ranking

• Link-graph influence: how much the shape of your links counts alongside the wording (0-1, default 1). This is the 3.0 channel described above; 0 reverts to ranking by content only. A live knob, no re-index.
• Idea influence: how much idea-level matching blends into the score (0-0.6). Notes are grouped into coherent ideas (~200-500 words); this weights whether two notes share a whole idea, not just one passage. 0 is passage-only. It is a live ranking knob; changing it re-ranks instantly with no re-index, so you can compare.
• Isolated areas: self-contained areas, one tag namespace per line (e.g. goa). A note tagged with an activated namespace (matching goa and goa/character) only relates to, and takes tag suggestions from, other notes in that area, and never appears in any other note's cards. Notes in no activated area share one pool. A live ranking knob (no re-index). Use it to keep a self-contained project (a novel, a world) from bleeding into unrelated notes.

Scope

• Excluded folders: folders left out of the index entirely (and everything beneath them); not ranked and not suggested as links. One per line or comma-separated.
• Folders excluded from link suggestions: folders whose notes stay indexed and ranked in the panel, but are never suggested as inline wikilinks.

Advanced

• Max chunks per note: ceiling on passages embedded per note. The whole note is covered up to this cap; only very long notes approach it.
• Heading context: embeds each section's first chunk with its note + heading breadcrumb for context. On by default; toggle to compare.
• Rebuild index: force a full re-embed (also on the command palette and the panel's refresh icon).

Changing the model or compute device transparently rebuilds the index; unrelated changes (sliders, toggles) never trigger a re-embed.

Install

Settings -> Community plugins -> Browse -> "Smart Related Notes" -> Install. That is the whole thing, and it is how almost everyone should install it.

On first launch the model weights download from the Hugging Face Hub with a progress notice, then cache. This happens once; after that the plugin works offline. The WASM (CPU) path, the default, uses quantized weights (~110 MB); the optional WebGPU path uses fp32 weights (~470 MB for the default model). The model is downloaded once per backend, then cached. The plugin also fetches the version-pinned ONNX runtime once, into its own ort/ folder.

Other ways

If the store has not caught up yet, or you want to track development, add this repository in BRAT, or download main.js, manifest.json and styles.css from the latest release into .obsidian/plugins/smart-related-notes/ yourself.

Requirements

• Desktop only (the embedding runtime needs a desktop Electron environment).
• Obsidian 1.7.2 or newer.

Development

A TypeScript project bundled with esbuild (entry src/main.ts → root main.js).

npm install          # install dev dependencies
npm run dev          # esbuild watch build (inline sourcemap, no minify)
npm run build        # gen-ort -> tsc --noEmit -> minified production bundle
npm run lint         # eslint (typescript-eslint + eslint-plugin-obsidianmd)

gen-ort.mjs runs before tsc/lint/esbuild: it writes src/ort-version.ts (the pinned onnxruntime-web version + the CDN base the runtime is downloaded from) and copies the matching onnxruntime-web .wasm/.mjs assets into ort/ for local dev builds. Both are build artifacts and are gitignored. Releases ship only main.js, manifest.json, and styles.css (exactly what Obsidian downloads); at runtime the plugin caches the pinned ONNX runtime into its ort/ folder on first use, validated against byte sizes pinned at build time so torn writes or a plugin update carrying a different runtime build trigger a clean re-download instead of mixing builds.

The embedding engine (transformers.js + onnxruntime-web) runs inside a dedicated Web Worker bundled as a second esbuild stage and inlined into main.js (see esbuild.config.mjs and src/worker/embed-worker.ts); terminating that worker is what returns the engine's memory to the OS when idle. Environments that expose a Node-like process would make transformers.js pick the (externalized, unavailable) onnxruntime-node backend — src/ort-shim.ts (imported first in the worker entry) flips the relevant process markers during module evaluation so the web runtime is used and WebGPU/WASM work.

License

MIT © 2026 Saiki77