A mature Obsidian vault rarely lacks information. The harder problem is retrieval: you remember the idea, but not the exact words or filename.
SystemSculpt combines exact vault search with embeddings-backed similarity so a query can find both literal matches and notes that express the same idea differently. The result is one search surface for project notes, transcripts, research, and drafts.
Why find notes by meaning
Keyword search is still best for titles, tags, identifiers, and phrases you know. Semantic search covers the opposite case: the decision was described as "runway pressure," but you remember "budget risk." Hybrid ranking keeps both kinds of evidence instead of forcing a choice between them.
That matters as vocabulary changes. A meeting, source PDF, and summary note can discuss the same subject without sharing the same wording. Meaning-based retrieval reconnects those records when memory is approximate.
The distinction is easiest to see with a real example. Imagine a vault that contains notes titled Q3 Cash Plan, Hiring Constraints, and Board Questions. None of them uses the phrase "budget risk" in the title. Exact search can still find useful matches if the phrase appears in the body, but it cannot infer that runway, hiring pace, and cash planning belong to the same decision. Semantic search gives those related concepts a chance to surface together.
That does not make exact search obsolete. If you know the project code, person's name, tag, or quoted phrase, exact matching is often faster and more precise. The useful setup combines both signals and lets the query determine which one matters more.
One SystemSculpt setup
SystemSculpt does not ask you to choose an embeddings provider or paste provider keys into Obsidian. Activate your SystemSculpt license, enable embeddings, and choose which vault folders should be excluded.
The client prepares note text, stores the current vector index locally, and watches for vault changes. Vector generation runs through the SystemSculpt API, so credentials, routing, and model changes stay off the client. Chat and Similar Notes use the same current index contract.
The managed AI guide explains which work happens through SystemSculpt's service and which data remains in the vault.
How indexing works
The first pass scans eligible Markdown files in bounded batches. Large vaults can take several minutes, but later updates are incremental. Excluding templates, archives, generated files, or private folders saves processing time and keeps irrelevant material out of results.
Search continues to work if embeddings are unavailable. SystemSculpt falls back to lexical scoring, so the feature degrades to exact vault search instead of becoming an empty screen.
If an index becomes unhealthy, run Rebuild SystemSculpt embeddings from the command palette. Duplicate-run safeguards prevent two rebuilds from racing over the same vault.
How to test search quality in your own vault
Do not judge a semantic search plugin from a demo vault. Test it against notes you already struggle to find.
I use a small retrieval check:
- Pick five notes you know well from different folders.
- Write one query using words that appear in each note.
- Write a second query that describes the idea without copying its language.
- Compare which notes appear and whether the top results are useful.
- Remove noisy folders, rebuild, and run the same queries again.
This separates search quality from the novelty of seeing embeddings run. A good result is not merely "the note appeared somewhere." The relevant note should appear early enough that you would actually use the search instead of falling back to folder browsing.
Also test failure behavior. Temporarily disconnect the network and confirm that exact vault search still returns useful results. A managed embeddings service can be unavailable; your core ability to find a filename or phrase should not disappear with it.
Privacy and indexing boundaries
Semantic search needs a representation of note content. In SystemSculpt, eligible note text is sent to the managed service to generate vectors, while the resulting index is stored in the vault environment. This is not an offline-only workflow.
Choose exclusions deliberately. Private journals, legal material, exported chat logs, generated archives, or folders with repeated templates may not belong in the index. Exclusions protect sensitive material and improve relevance by removing content that would otherwise compete with the notes you care about.
The right boundary is specific to the vault. A research library may benefit from broad indexing. A mixed personal and work vault may need tighter folder rules. Review those rules whenever the vault structure changes.
From retrieval to action
Finding a note is usually the beginning of the task. A retrieved audio transcript can become context for a project brief. Related research notes can feed a comparison. Old decisions can be cited before an agent proposes an update.
SystemSculpt keeps that next step visible. Read-only tools can gather context immediately; file-changing tools show the proposed path or diff and follow the chat's approval mode before touching the vault.
Get started
Enable embeddings under Settings > SystemSculpt > Embeddings, set exclusions, and let the first index finish. Then open SystemSculpt Search and try queries based on ideas you know exist but cannot locate by phrase.
The current setup and troubleshooting path is in the embeddings and semantic search documentation. You can also review the full SystemSculpt plugin surface before deciding whether managed semantic search fits your vault.