Sourcebound is a local-first document search workspace. It helps people retrieve exact passages from their own notes without uploading private source material to a third party.
Personal knowledge is spread across files and difficult to retrieve. Cloud AI tools can answer questions, but often obscure which passage supports an answer and require users to surrender their documents. The first product risk is retrieval quality, not text generation. The MVP therefore proves import, indexing, search, and source inspection.
Researchers, developers, students, and writers with a folder of Markdown or text notes who need to find a remembered fact and verify it against the original wording.
A user can build and search a private local library in under two minutes, then open the complete source behind any matching passage.
- As a user, I can import a
.mdor.txtfile up to 5 MB. - As a user, I can see every imported document in a library.
- As a user, I can search all document contents and see highlighted excerpts.
- As a user, I can open the complete original source from a result.
- As a user, I can remove a document and its search index data.
- As a privacy-conscious user, I can use all MVP features without an external service.
- Imported documents persist in a local SQLite database across restarts.
- Search results appear within 500 ms for a library of 1,000 typical notes.
- A result displays its filename and a passage containing highlighted matching terms.
- Search syntax failures return actionable feedback without crashing the app.
- Unsupported and empty files are rejected.
- Deleting a document removes it from the library and subsequent search results.
- Desktop and mobile layouts remain usable down to a 320 px viewport.
- Core indexing and deletion behavior is covered by automated tests.
- PDF, DOCX, image, or OCR ingestion
- Accounts, synchronization, collaboration, or hosted storage
- Embedding-based semantic search
- LLM-generated answers
- Rich Markdown rendering or editing
- Folder watching and bulk import
- Native desktop packaging
- At least 80% of five pilot users can import and find a supplied passage unassisted.
- Zero document contents leave the machine during the core workflow.
- Search returns the expected source in the top five for 90% of a 50-query test set.
- No data loss across restart, deletion, and failed-import test scenarios.
| Risk | MVP mitigation |
|---|---|
| FTS syntax surprises users | Debounced search and actionable errors |
| Binary or oversized files exhaust memory | Extension allowlist and 5 MB limit |
| “Local-first” is confused with browser-only | State that the local API and SQLite file remain on-device |
| Generated answers distract from weak retrieval | Defer generation until retrieval evaluation passes |
MVP is complete when all acceptance criteria pass locally and in CI, the repository has public contribution documentation, and one pilot user completes the import-search-open workflow.