Ask your notes.
Open the sources.
Ask your Obsidian notes, then open the original passages behind an answer. Start with three sample notes and a ready answer connection. Explore connected local knowledge, a knowledge graph and insights you can trace; optional Jev adds ontology-style organization.
0.4.8 alpha. Desktop Obsidian 1.11.4+. Available through the official community listing and Release 0.4.8.
Optional support. No feature unlocks.Build knowledge
you can ask.
Notework uses retrieval-augmented generation (RAG): local retrieval finds passages in the notes you choose. Saved AI relationship judgments can bring related notes into retrieval. Optional ontology-style Build analysis adds category, role, overview/topic/detail and candidate-relationship judgments to help you organize and explore that knowledge.
Categories group topics; roles describe a note’s purpose; hierarchy marks overview, topic and detail levels; candidate relationships suggest connections between sources. Optional Jev analysis adds these judgments at Build. Check them against the original notes.
- Choose your local knowledge.Set folders, tags and exclusions. Use local Ollama embeddings or explicit keyword search to retrieve passages.
- Add structure when useful.Your Jev key enables optional category, role, hierarchy and candidate-relationship judgments at Build. Selected-AI analysis is another build route; local retrieval works without Jev.
- Ask with evidence for this question.Reference notes, retrieved excerpts and graph highlights refresh for each question. Inspect retrieval details and open the original passages.
- Follow connections toward insight.Explore related notes, compare their evidence and refine your question. Use the generated interpretation alongside the source material.
Answer generation, local retrieval and optional Build analysis are separate choices. Choose the routes that fit your notes and review which excerpts each route sends.
Ask with your local knowledge.
Ask about the collection you choose. Notework retrieves relevant Markdown passages locally, then gives your selected answer model the question and eligible evidence.
The right sidebar keeps your current note in view. Read Markdown answers, inspect the reference notes and retrieval details that refresh for each question, and open the original passages when a claim matters.
- Scope your context with folders, tags and explicit exclusions.
- Keep model and reasoning choices close to the question.
- Save conversations and reuse prompts as ordinary vault notes.
Source links help you check an answer. They do not prove that every generated statement is supported.
Explore your knowledge graph.
Small, bright note stars sit in a dark, navigable 3D space. Stable topic colors, hierarchy and role views help you explore a collection without losing the conversation.
Zoom toward the cursor, rotate the graph, and select a star to open its original note. Source highlights connect the current answer with the notes it used. Follow those notes to compare evidence and develop an insight you can trace back to its sources.
Graph filters change what you see. Use Pin selected context when you want to narrow future retrieval. Star positions are a navigation layout, not a measured similarity distance.
Keep the thread of a conversation.
After a completed answer, your selected AI automatically organizes exchanges into topics, refinements, branches and follow-ups. The summary dock helps you revisit a question without replacing it with another summary.
Open a card to read its exact original question and answer. Move between the map and the existing chat while your graph and document stay open.
Every completed answer adds one bounded structure request to the same selected answer model: up to 20 eligible pairs within 32 KiB of serialized text. Eligibility verification covers at most 64 distinct source references across that window. The same provider limits or API billing apply. Exchanges outside that verified window remain readable locally. Opening the dock makes no request.
Look past the answer.
The useful next step is often reading the passage behind a claim. Source evidence makes that path short.
Try the authored examples below. This is an illustration of evidence inspection, not a live plugin or AI connection.
Illustration · invented notes
Research / Review checklist.md
Review checklist
“Check the original passage for every key claim before sharing the summary.”
A source link gives you a place to inspect. Compare the answer with the full note, including its surrounding context.
Start with three notes.
Check one answer.
Open the official community listing and choose Add to Obsidian. For manual installation, download version-matched main.js, manifest.json and styles.css from Release 0.4.8 into your vault's Notework plugin folder.
This exercise uses fictional notes and a ready answer connection. Keyword search and local structure need no embedding-model download or Jev key. Sending a question still uses your chosen AI connection and its usage terms.
Follow the first-question guide- Connect your answer model.Keep your ready connection, or use Setup to connect official Codex or Claude Code, your own API key, or local Ollama chat. Connected does not guarantee remaining allowance.
- Copy three fictional notes.Save the Harbor pilot sample files in a vault folder named Notework First Question. Select only that folder in Scope and choose Apply scope.
- Build the small starting scope.In Build settings choose Keyword search and Local structure only, then Build knowledge. Wait for saving to finish. Jev analysis is optional and can wait.
- Ask your first question.Open Notework and ask: “What is the Harbor pilot review date, why did it change, and what must be ready before the review? Cite the notes and say what is still undecided.”
- Open the originals.Expand Source notes, inspect retrieved excerpts and open the note paths. Compare the old review date with the later decision and check the remaining tasks. These are authored source checks, not a recorded AI result.
Your AI.
Clear boundaries.
The MIT-licensed plugin has no access fee. Accounts, API usage and models have their own terms. Choose answer generation, retrieval embeddings and optional knowledge analysis separately.
A Connected indicator confirms authentication/catalog readiness. It does not guarantee remaining allowance, every model, or successful inference.
| Connection | What you bring | Usage and costs |
|---|---|---|
| ChatGPT / Codex | Separately installed official Codex app-server and your own official login. | Your account's available models and Codex limits. No automatic API fallback. |
| Claude Code | Separately installed, unmodified official CLI and direct user login. | Eligible account/subscription limits. A discovered model is not proof of entitlement. |
| OpenAI / Anthropic API | Your own provider API key. | API billing is separate from subscriptions. |
| Ollama chat & embeddings | A running local server and separate installed chat/embedding models. | No per-call cloud API charge for local inference. Your hardware, model license and server configuration matter. |
| Optional Jev analysis | Your own Jev API key and available model. | Separate usage for optional ontology-style category, role, hierarchy and candidate-relationship judgments at Build. Local retrieval works without it. |
Automatic conversation structure uses the selected answer model and adds usage after an answer. Local embeddings do not pay for cloud answers, and Jev does not structure the conversation.
Know where your notes go.
Local retrieval gives you a useful boundary, but cloud answers still send data to the connection you choose. Review your scope before using sensitive material.
Cloud questions send selected evidence.
Your question, eligible history, retrieved excerpts and note paths, and explicit text attachments go to the selected provider. Automatic structure sends eligible completed exchanges. Optional AI/Jev builds send selected note excerpts.
Local routes and records stay local.
Ollama requests use the local loopback server. Indexes contain unencrypted excerpts, vectors and source metadata. Optional Markdown conversation archives and prompt notes can sync with your vault.
Desktop connections use installed tools.
Notework discovers official CLIs through PATH/configured paths, checks installation metadata, starts processes, and creates dedicated system-temporary directories outside the vault. It does not install these tools. Restricted execution controls are not an OS sandbox.
Credentials have a defined storage boundary.
API keys and retained direct-login credentials use vault-local Obsidian SecretStorage. It is not an OS keychain or isolation from other plugins. Official Codex and Claude Code retain their own login. Provider retention policies apply independently.
No developer analytics endpoint or automatic diagnostic upload was found in the reviewed implementation. Authentication, model/catalog requests and explicit downloads still use network connections.
Read the full security and data-handling policyA few useful answers.
Can I use my ChatGPT or Claude subscription?
The desktop connection paths use the separately installed official Codex app-server or unmodified Claude Code CLI, with your own direct login. Your account's eligibility, model access and usage limits apply. API connections are separate and use API billing. Notework does not promise every plan or every model.
Do I need Jev, or a paid cloud embedding API?
Neither is required for local retrieval. Choose local Ollama embeddings or explicit keyword search. Jev is optional analysis for categories, roles, hierarchy and candidate relationships during Build. It uses your own key and provider allowance. The answer model is a separate connection.
How do RAG, ontology and the knowledge graph fit together?
RAG retrieves passages from your selected local notes for the answer model. Related-note retrieval uses saved AI relationship judgments that remain valid for the selected scope. The knowledge graph lets you navigate the collection and inspect source notes. Optional ontology-style Build analysis adds category, role, overview/topic/detail and candidate-relationship judgments, with Jev as one optional analysis route. Inspect these model judgments and develop insights by comparing evidence and following connections.
Does the conversation map cost extra?
It adds an automatic request after each completed answer under the same selected answer model's limits or API billing. The request is bounded to up to 20 eligible pairs within 32 KiB of serialized UTF-8 text. Eligibility verification covers at most 64 distinct source references across that window. Exchanges outside that verified window remain local. Opening or navigating the summary makes no request.
Can I install it from Community Plugins?
Yes. Open the official community listing and choose Add to Obsidian. You can also install the version-matched main.js, manifest.json and styles.css from Release 0.4.8 manually.
Does it work on my phone?
Mobile is under active development and currently disabled. This release is desktop-only. Disable an older phone copy on the phone itself and keep active/installed community-plugin-list Sync off on both devices. Ordinary note Sync can continue.
Are these screenshots real vault data?
The desktop captures use invented notes and synthetic provider state. The star graph at the top and interactive source examples are authored illustrations. They show the product's working idea and interface, not paid provider results, account access or private vault content.
Where are the website and source?
The website and plugin source use one repository. GitHub Pages serves its docs folder at hpsonghyun.github.io/notework-ai. The older personal-site repository is not a second development target.