obsearch — Obsidian Multimodal Vault Search

Monorepo for indexing an Obsidian vault and searching across text, images, and PDFs using Gemini embeddings + local SQLite vector search.

Current Status (2026-03-11)

Fase 1 — MVP

1. Monorepo (Bun workspaces + Turborepo)
2. @obsearch/core Gemini embedding + SQLite/sqlite-vec schema
3. Vault crawler (.md, images, .pdf)
4. TaskType embedding API (RETRIEVAL_DOCUMENT / RETRIEVAL_QUERY)
5. Image indexing pipeline (indexImage)
6. Markdown indexing pipeline (indexMarkdown)
7. Retry + exponential backoff (retryWithBackoff)
8. Incremental indexing by mtime + size (indexVaultFile)
9. CLI index <vault-path> (apps/cli)
10. oRPC search(query, limit) endpoint
11. Web UI with thumbnails + Obsidian deep links

Fase 2

• Semantic chunking for long markdown
• PDF text extraction + chunking pipeline
• Snippet highlight in search results

Implemented Core APIs (@obsearch/core)

• crawlVault({ vaultPath })
• createEmbeddingClient({ expectedDimension: 3072 })
• initDb({ dbPath, embeddingDimension })
• indexImage(...)
• indexMarkdown(...)
• indexVaultFile(...) (incremental: indexed / skipped with reason)
• retryWithBackoff(...)
• search(query, limit) on CoreDb

CLI (apps/cli)

Run index command

bun run apps/cli/src/index.ts index /absolute/path/to/vault

or from the CLI package:

bun run --cwd apps/cli index -- /absolute/path/to/vault

CLI behavior

• Crawls supported files from vault
• Initializes DB at <vault>/.obsearch/index.db
• Indexes per file with progress output [i/total]
• Uses incremental skip logic (mtime + size)
• Prints summary: indexed, skipped, errors, duration
• Prints heuristic API cost estimate

Cost estimate formula:

embedding_call_count * OBSEARCH_ESTIMATED_USD_PER_EMBED_CALL

If OBSEARCH_ESTIMATED_USD_PER_EMBED_CALL is unset, default is 0.

API + Web Status

• API route implemented: search({ query, limit? }) in @obsearch/api.
• Server now wires a real coreDb.search(...) dependency into API context.
• Web home route includes:
  • Search form (query, optional limit)
  • Result cards with type + score
  • Real image thumbnails via server route /vault-file/<relative-path>
  • Obsidian deep links per result

Environment Variables

CORS_ORIGIN=http://localhost:3001
GEMINI_API_KEY=your_key_here
GEMINI_EMBEDDING_MODEL=gemini-embedding-2-preview
OBSEARCH_DB_PATH=/absolute/path/to/vault/.obsearch/index.db
OBSEARCH_VAULT_PATH=/absolute/path/to/vault
OBSEARCH_THUMBNAIL_TOKEN=change_me
VITE_OBSIDIAN_VAULT_NAME=YourVaultName
VITE_OBSEARCH_THUMBNAIL_TOKEN=change_me
OBSEARCH_ESTIMATED_USD_PER_EMBED_CALL=0.00010

Development

bun install
bun run dev

Useful checks:

• bun run --cwd packages/core test
• bun run --cwd packages/core check-types
• bun test apps/cli/src/index.test.ts
• bun run --cwd apps/cli check-types