[Leo310]

Found during the live test of #490: after cancelling a build on a newly added index and reloading Obsidian, every note in the resumed run failed with Node with id N already exists.

Root cause

The live database had 450 graph nodes, 537 rows, a graph header, and no metadata record. Without it HNSWVectorStore.open left the id counter at 0, so the resumed run reallocated ids into the existing graph. Each failed insert had already overwritten the id mapping of the note that owned that id, so the index is corrupted beyond what a resume can repair.

Which run writes rows without metadata? The startup validation. "Add index" calls ensureIndex, which opens the instance; opening schedules validation on a zero-delay timer, and that timer fires before ensureIndex gets through its getMetadata and count worker round-trips. Validation sees an empty store, embeds everything through embedFilesInBatches, and the full build that ensureIndex then attempts bails with "Indexing already in progress". Only the full build calls setMetadata, and updateLastUpdated has no record to persist the counter into. The index also never gets a lastBuiltAt.

What changed

• Store: open sets the counter to the max of the metadata high-water mark and the loaded id mappings, so a missing or stale record can never reset it below an id in use.
• Service: validation on an empty store with vault files runs buildFullIndex instead of a repair. buildFullIndex keeps one build promise per instance; a second caller joins it, so the modal's ensureIndex and the scheduled validation no longer race or notice against each other.
• Repair: validation writes the metadata record when a store has rows but none, so indexes built through the old path restore dimensions and counter normally on the next open.

Tests

• HNSWVectorStore.resume.test.ts: rows and graph written with no metadata record, reopen, upsert succeeds and both old and new notes resolve to their own rows.
• VectorStoreService.lazyOpen.test.ts: with the store's reads parked on a macrotask so validation fires first, a fresh index gets exactly one build, the metadata record exists, and lastBuiltAt is stamped.

Both fail on main and pass here. bun run check, format, lint, test (1677 passing) and a dev build are clean; the bundle loads in a live vault without errors.

Note for anyone who hit this: an index that already collided needs Rebuild index; the overwritten mappings cannot be reconstructed.

[greptile-apps]

Greptile Summary

This PR prevents interrupted or startup-raced vector index builds from reusing persisted HNSW node IDs and coordinates concurrent requests for the same full build.

• Restores the ID counter from metadata, ID mappings, and persisted graph nodes.
• Routes validation of a fresh store through the full-build lifecycle.
• Makes concurrent full-build callers share one per-instance promise.
• Repairs missing metadata for nonempty stores, including vector dimensions.
• Adds regression coverage for missing metadata, interrupted removal, and startup timing.

Confidence Score: 5/5

The PR appears safe to merge; no actionable correctness, security, or repository-rule issue remains.

Both previous findings are fully addressed in the current code: ID restoration includes persisted graph keys, and metadata repair runs for every nonempty metadata-less store. The new build promise also makes the startup validator and ensureIndex await the same full build.

Important Files Changed

Filename	Overview
src/vectorstore/HNSWVectorStore.ts	Restores the HNSW ID high-water mark from every persisted numeric-ID source and probes existing row dimensions during metadata repair.
src/vectorstore/VectorStoreService.ts	Sends empty-store validation through a deduplicated full build and repairs missing metadata independently of embedding work.
test/vectorstore/HNSWVectorStore.resume.test.ts	Covers reopening metadata-less stores and retaining graph IDs whose mappings were interrupted during removal.
test/vectorstore/VectorStoreService.lazyOpen.test.ts	Covers startup validation winning the scheduling race while still producing exactly one complete build.

Sequence Diagram

sequenceDiagram
    participant E as ensureIndex
    participant V as Startup validation
    participant B as buildFullIndex
    participant S as Vector store

    V->>S: Inspect rows and metadata
    E->>S: Inspect metadata and count
    V->>B: Request full build for empty store
    B->>B: Store buildPromise
    E->>B: Request same full build
    B-->>E: Return existing buildPromise
    B->>S: Embed rows, flush graph, write metadata
    B-->>V: Build completes
    B-->>E: Shared build completes

Loading

_{Reviews (3): Last reviewed commit: "fix(vectorstore): count persisted graph ..." | Re-trigger Greptile}