fix(hnsw): skip unvectored rows on rebuild; refuse empty vectors in the index
Some checks failed
CI / Node 22 (push) Successful in 12m14s
CI / Node 24 (push) Successful in 12m15s
CI / Integration + conformance (Node 22) (push) Failing after 14m55s
CI / Bun (latest) (push) Successful in 12m24s

A canonical row persisted with vector: [] (a system row, a deferred embed
not yet landed, or any other legitimately-unvectored record) is a normal,
enumerable row -- but rebuild()'s storage walk had no guard against it.
storage.getVectorIndexData() derives its answer from the row's own record,
so it returns non-null for any existing noun whether or not that noun was
ever actually indexed -- rebuild() admitted such rows into the live graph
with a length-0 vector. A vector-less node could become the entry point (or
occupy any graph position); the next real insert then ran a distance
calculation against it and blew up with a dimension mismatch.

Fix at two layers in src/hnsw/hnswIndex.ts:
- rebuild() now skips any row whose vector.length === 0 before it ever
  becomes a graph node (one summary count line, never per-row spam), and
  restores the pinned dimension from the first real vector it loads --
  previously the pin stayed null across a restart, since addItem/updateItem
  are the only sites that set it and rebuild() never goes through either.
- addItem/updateItem now refuse a length-0 vector with a typed
  EmptyVectorIndexError instead of ever pinning dimension to 0 or storing a
  vector-less node, so no future fill/rebuild/load path can poison the index
  silently. getVectorSafe's lazy-load "not found" check also missed that an
  empty array is truthy -- tightened to catch it.

IndexOperations.ts's ReplaceInVectorIndexOperation rollback paths now skip
re-adding an oldVector of length 0 (never a legal index member) instead of
attempting an illegal empty re-insert on rollback.

biography.test.ts's final ledger-exactness assertion assumed every noun the
lane creates is vectored, including the VFS root counted in
vfsBaselineNouns -- but the root is deliberately persisted unvectored.
Corrected the expected formula to exclude it.

Adds tests/integration/index-skips-unvectored.test.ts pinning: rebuild()
indexes only vectored rows with the dimension pinned correctly; clear()
then real adds never trip a dimension mismatch; addItem/updateItem refuse a
length-0 vector; and a crash/repair cycle stays dimension-consistent.
This commit is contained in:
David Snelling 2026-08-27 13:29:11 -07:00
parent fd6b4ce4ff
commit 8fc553b126
4 changed files with 371 additions and 8 deletions

View file

@ -64,6 +64,34 @@ export class HnswFlushError extends Error {
}
}
/**
* @description Thrown by {@link JsHnswVectorIndex.addItem} / {@link
* JsHnswVectorIndex.updateItem} when handed a length-0 vector. A length-0
* vector is the sanctioned "unvectored" shape for a canonical noun record
* (class-J: a VFS-system row, a deferred embed not yet landed, or any other
* legitimately-vector-less row) but it is NEVER a legal INDEX insert. The
* index itself has no concept of "unvectored"; deciding that a row is
* unvectored and therefore skippable is the FILL/REBUILD/LOAD consumer's job
* (see {@link JsHnswVectorIndex.rebuild}), done BEFORE ever calling addItem.
* A length-0 vector reaching this point is a caller bug: silently accepting
* it would pin `this.dimension = 0` on an empty index (poisoning every real
* insert thereafter with a dimension mismatch) or store a vector-less node
* that a distance calculation can never safely compare against. Loud errors,
* never quiet losses this throws instead of either.
*/
export class EmptyVectorIndexError extends Error {
constructor(public readonly id: string, operation: 'addItem' | 'updateItem') {
super(
`${operation}(${id}): refusing to index a length-0 vector — a length-0 vector is the ` +
`sanctioned "unvectored" shape for a canonical row, but it is never a legal index ` +
`insert. Callers that fill/rebuild/load the index must skip vector.length === 0 rows ` +
`themselves (unvectored = nothing to index, not an error at that layer); reaching ` +
`here with one is a caller bug.`
)
this.name = 'EmptyVectorIndexError'
}
}
/**
* Implements {@link VectorIndexProvider}: the vector-index surface Brainy calls
* on whatever the `'vector'` factory returns (its own `JsHnswVectorIndex`, or a native
@ -580,6 +608,15 @@ export class JsHnswVectorIndex implements VectorIndexProvider {
throw new Error('Vector is undefined or null')
}
// THE INDEX REFUSES A LENGTH-0 VECTOR (see EmptyVectorIndexError's JSDoc):
// an empty vector is the sanctioned "unvectored" shape at the canonical
// layer, never a legal index member. Refusing here — loudly, before the
// dimension pin below — means no future fill/rebuild/load path can ever
// poison `this.dimension` to 0 or park a vector-less node in the graph.
if (vector.length === 0) {
throw new EmptyVectorIndexError(id, 'addItem')
}
// Set dimension on first insert
if (this.dimension === null) {
this.dimension = vector.length
@ -954,6 +991,13 @@ export class JsHnswVectorIndex implements VectorIndexProvider {
return
}
// Same refusal as addItem (see EmptyVectorIndexError's JSDoc) — an
// in-place relink must never rewrite an already-indexed node down to the
// unvectored shape or poison the pinned dimension.
if (vector.length === 0) {
throw new EmptyVectorIndexError(id, 'updateItem')
}
if (this.dimension === null) {
this.dimension = vector.length
} else if (vector.length !== this.dimension) {
@ -1555,7 +1599,15 @@ export class JsHnswVectorIndex implements VectorIndexProvider {
}
const loaded = await this.storage.getNounVector(noun.id)
if (!loaded) {
// `loaded` is a length-0 array (not null/undefined) for a canonical row
// that is legitimately unvectored — `![]` is FALSE (an empty array is
// truthy), so the bare `!loaded` check below would silently accept it
// as "found" and hand a dimension-0 vector to a distance calculation.
// A node only reaches this lazy-load path because it is a MEMBER of
// the live index (rebuild() now refuses to admit unvectored rows — see
// its JSDoc), so an empty vector here is never legitimate: treat it
// exactly like "not found", loudly.
if (!loaded || loaded.length === 0) {
throw new Error(`Vector not found for noun ${noun.id}`)
}
@ -1765,9 +1817,42 @@ export class JsHnswVectorIndex implements VectorIndexProvider {
totalCount = result.totalCount || result.items.length
// UNVECTORED ROWS ARE NOT AN INDEX MEMBER (the class-J law): a canonical
// noun whose vector leg is `[]` (a VFS-root-style system row, a
// deferred embed not yet landed, or a best-effort fallback for an
// unreadable vector leg) is a normal, enumerable, countable row — it
// is simply not indexed. `storage.getVectorIndexData()` derives its
// {level, connections} answer straight from the noun's OWN record, so
// it returns non-null for every existing noun regardless of whether
// that noun ever actually reached `addItem()` — it cannot be used to
// decide indexability. `nounData.vector.length === 0` is the one
// truthful signal (mirrors the `noun.vector.length > 0` guards in
// {@link getVectorSafe} / {@link getVectorSync}): skip here, counted
// once in a summary line, never per-row spam.
let skippedUnvectored = 0
// Process all nouns at once
for (const nounData of result.items) {
try {
if (!nounData.vector || nounData.vector.length === 0) {
skippedUnvectored++
continue
}
// Restore the pinned dimension from the first real vector this
// rebuild loads. `addItem`/`updateItem` only pin `this.dimension`
// on a LIVE insert — a fresh rebuild from storage never goes
// through either, so without this the pin stays `null` across a
// restart. A `null` pin means the very next insert (correct OR
// wrong length) silently BECOMES the new pin instead of being
// checked against the store's real dimension — the wrong-length
// case then fails much later and less clearly, inside a distance
// calculation against an already-loaded node, instead of here,
// immediately, with a named expected-vs-got mismatch.
if (this.dimension === null) {
this.dimension = nounData.vector.length
}
// Load HNSW graph data for this entity
const hnswData = await this.storage.getVectorIndexData(nounData.id)
@ -1815,7 +1900,10 @@ export class JsHnswVectorIndex implements VectorIndexProvider {
options.onProgress(loadedCount, totalCount)
}
prodLog.info(`HNSW: Loaded ${loadedCount.toLocaleString()} nodes (${storageType})`)
prodLog.info(
`HNSW: Loaded ${loadedCount.toLocaleString()} nodes (${storageType})` +
(skippedUnvectored > 0 ? `${skippedUnvectored.toLocaleString()} unvectored row(s) skipped` : '')
)
}
// Step 5: CRITICAL - Recover entry point if missing)

View file

@ -324,8 +324,17 @@ export class ReplaceInVectorIndexOperation implements Operation {
return async () => {
// Restore the declared before-state in place (see class JSDoc for
// the item-did-not-exist posture).
// the item-did-not-exist posture). A length-0 oldVector means the row
// was never actually indexed before this op ran (a length-0 vector is
// never a legal index member — see EmptyVectorIndexError) — there is
// no in-place "restore to empty" for the provider to perform, so
// rollback removes the row instead, leaving the same "not indexed"
// state the row was in before execute().
if (this.oldVector.length > 0) {
await index.updateItem!({ id: this.id, vector: this.oldVector }, generation)
} else {
await this.index.removeItem(this.id, generation)
}
}
}
@ -336,11 +345,16 @@ export class ReplaceInVectorIndexOperation implements Operation {
return async () => {
// updateItem-style restore via the same adjacent pair, back to the
// declared before-state.
// declared before-state. Same length-0 carve-out as the updateItem
// path above: an empty oldVector was never a legal index member, so
// rollback just leaves the row removed rather than attempting an
// illegal empty re-add.
await this.index.removeItem(this.id, generation)
if (this.oldVector.length > 0) {
await this.index.addItem({ id: this.id, vector: this.oldVector }, generation)
}
}
}
}
/**

View file

@ -0,0 +1,253 @@
/**
* @module tests/integration/index-skips-unvectored
* @description THE UNVECTORED-ROW CURE two integration tests
* (`tests/lifecycle/biography.test.ts`'s Ch4/5/6 chapter and
* `tests/integration/clear-persistence.test.ts`'s multi-cycle test) started
* failing after a canonical-storage change made a vector-less row (the
* class-J shape: `vector: []`, e.g. the VFS root, a deferred embed not yet
* landed, or any other legitimately-unvectored canonical record) VISIBLE to
* the enumeration walk `getNounsWithPagination()` for the first time before
* that change such rows were simply invisible to the walk. `hnswIndex.ts`'s
* `rebuild()` never guarded against that shape: it inserted every row the
* walk yielded into the live in-memory index, including ones with a length-0
* vector, because `storage.getVectorIndexData()` derives its {level,
* connections} answer straight from the noun's OWN record it returns
* non-null for ANY existing noun, whether or not that noun was ever actually
* indexed via `addItem()`. A vector-less node admitted into the graph could
* become the entry point (or occupy any graph position), and the very next
* real-vectored `addItem()` then ran a distance calculation against it
* `cosineDistance` throws "Vectors must have the same dimensions" the moment
* one operand is a length-0 array.
*
* THE FIX, at two layers (`src/hnsw/hnswIndex.ts`):
* (1) FILL/REBUILD/LOAD consumers treat `vector.length === 0` as "unvectored
* nothing to index" and skip the row (normal, not an error; one summary
* count line, never per-row spam) `rebuild()`'s loop now checks this
* BEFORE ever creating a graph node, so an unvectored row can never
* become an index member, entry point, or dimension-setter.
* (2) THE INDEX ITSELF refuses a length-0 vector in `addItem()` /
* `updateItem()` with a typed `EmptyVectorIndexError`, loudly, instead of
* ever pinning `dimension = 0` or storing a vector-less node so no
* future fill/rebuild/load path can silently poison the index even if it
* forgets law (1).
*
* Four legs pinned here:
* (a) `rebuild()` over a store mixing real-vectored rows and `vector: []`
* rows indexes ONLY the vectored ones size === vectored count,
* dimension pinned to the real (non-zero) length.
* (b) `clear()` then real adds afterward never trip a dimension mismatch
* the exact `clear-persistence.test.ts` regression shape, reproduced
* directly against the index/storage seam this module owns.
* (c) `index.addItem({ id, vector: [] })` throws `EmptyVectorIndexError`
* (and `updateItem` does too, for an existing node).
* (d) crash -> repair: the crashed generation's entities survive, the ledger
* recounts honestly, and a fresh real-vectored add afterward never trips
* a dimension mismatch against a leftover vector-less phantom.
*/
import { describe, it, expect, afterEach } from 'vitest'
import * as fs from 'node:fs'
import * as os from 'node:os'
import * as path from 'node:path'
import { Brainy } from '../../src/index.js'
import { NounType } from '../../src/types/graphTypes.js'
import { EmptyVectorIndexError } from '../../src/hnsw/hnswIndex.js'
import { abandonAsCrashed, openBrain as openKillMatrixBrain, uid, vec } from '../helpers/durabilityKillMatrix.js'
const tmpDirs: string[] = []
function mkTmp(): string {
const d = fs.mkdtempSync(path.join(os.tmpdir(), 'brainy-index-skips-unvectored-'))
tmpDirs.push(d)
return d
}
afterEach(() => {
for (const d of tmpDirs.splice(0)) {
try {
fs.rmSync(d, { recursive: true, force: true })
} catch {
/* best-effort cleanup */
}
}
})
/** A filesystem-backed brain with explicit vectors (no embedder needed) and
* manual persistence mirrors `durabilityKillMatrix.ts`'s `openBrain` so
* every write in this module is explicit and provably durable. */
function openBrain(dir: string): any {
return new Brainy({
requireSubtype: false,
storage: { type: 'filesystem', path: dir },
silent: true,
persistence: { policy: 'manual' }
})
}
describe('HNSW index skips unvectored rows', () => {
it('(a) rebuild() indexes only vectored rows: size === vectored count, dimension pinned to the real length', async () => {
const dir = mkTmp()
let brain = openBrain(dir)
await brain.init()
// 5 real-vectored rows.
const vectoredIds: string[] = []
for (let i = 0; i < 5; i++) {
const id = uid(`vectored-${i}`)
await brain.add({ id, data: `real entity ${i}`, type: NounType.Document, vector: vec(i) })
vectoredIds.push(id)
}
// 3 explicit unvectored rows — the class-J "vector: []" shape, a normal,
// enumerable, countable canonical row that must never reach the index.
const unvectoredIds: string[] = []
for (let i = 0; i < 3; i++) {
const id = uid(`unvectored-${i}`)
await brain.add({ id, data: `unvectored entity ${i}`, type: NounType.Document, vector: [] })
unvectoredIds.push(id)
}
await brain.flush()
// The canonical ledger already agrees before any rebuild: nouns.all
// counts every row (8 + the VFS root); vectors.all counts only the real
// ones (5) — the VFS root and the 3 explicit unvectored rows are excluded.
const ledgerBeforeReopen = await brain.storage.getCanonicalCounts()
expect(ledgerBeforeReopen.vectors.all).toBe(5)
expect(ledgerBeforeReopen.nouns.all).toBe(9) // 5 vectored + 3 unvectored + 1 VFS root
await brain.close()
// Reopen: open()'s index build IS hnswIndex.rebuild() run fresh from
// storage — this is the exact path that used to admit unvectored rows.
brain = openBrain(dir)
await brain.init()
const status = await brain.getIndexStatus()
expect(status.hnswIndex.size, 'the rebuilt index must contain ONLY the 5 real-vectored rows').toBe(5)
// Dimension is pinned to the REAL embedded length (384 via `vec()`), not
// 0 — adding a wrong-length vector must be refused naming that real
// dimension, proving no vector-less row ever set it.
const realDimension = vec(0).length
let mismatchMessage: string | undefined
try {
await brain.index.addItem({ id: uid('dimension-probe'), vector: vec(0).slice(0, realDimension - 1) })
expect.fail('expected a dimension mismatch error')
} catch (err) {
mismatchMessage = (err as Error).message
}
expect(mismatchMessage).toContain(`expected ${realDimension}`)
// Every unvectored row is still a normal, enumerable, readable canonical
// record — class-J semantics survive the rebuild fix untouched.
for (const id of unvectoredIds) {
const entity = await brain.get(id, { includeVectors: true })
expect(entity, `unvectored entity ${id} must remain readable`).not.toBeNull()
expect(entity.vector).toEqual([])
}
// A correct-dimension add succeeds cleanly against the pinned dimension.
const freshId = uid('post-reopen-fresh')
await expect(brain.add({ id: freshId, data: 'fresh', type: NounType.Document, vector: vec(50) })).resolves.toBe(
freshId
)
await brain.close()
})
it('(b) clear() then real adds afterward never trip a dimension mismatch (the clear-persistence regression shape)', async () => {
const dir = mkTmp()
let brain = openBrain(dir)
await brain.init() // the VFS root (vector: []) is the store's only row
await brain.clear()
await brain.close()
// Reopen over a store whose only surviving row is the recreated,
// unvectored VFS root — this is exactly the shape that used to poison
// the entry point / dimension in `clear-persistence.test.ts`.
brain = openBrain(dir)
await brain.init()
expect((await brain.getIndexStatus()).hnswIndex.size).toBe(0)
const id1 = uid('after-clear-1')
await expect(brain.add({ id: id1, data: 'after clear 1', type: NounType.Document, vector: vec(1) })).resolves.toBe(
id1
)
const id2 = uid('after-clear-2')
await expect(brain.add({ id: id2, data: 'after clear 2', type: NounType.Document, vector: vec(2) })).resolves.toBe(
id2
)
expect((await brain.getIndexStatus()).hnswIndex.size).toBe(2)
await brain.close()
})
it('(c) index.addItem/updateItem refuse a length-0 vector with EmptyVectorIndexError', async () => {
const dir = mkTmp()
const brain = openBrain(dir)
await brain.init()
await expect(brain.index.addItem({ id: uid('empty-add'), vector: [] })).rejects.toThrow(EmptyVectorIndexError)
// updateItem on an EXISTING (real-vectored) node must refuse the same way.
const existingId = uid('existing-for-update')
await brain.add({ id: existingId, data: 'existing', type: NounType.Document, vector: vec(9) })
await expect(brain.index.updateItem({ id: existingId, vector: [] })).rejects.toThrow(EmptyVectorIndexError)
// The index was never disturbed by either refused call.
expect((await brain.getIndexStatus()).hnswIndex.size).toBe(1)
await brain.close()
})
it('(d) crash -> repair: the crashed generation survives, the ledger recounts honestly, and a fresh add afterward never trips a dimension mismatch', async () => {
const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'brainy-index-skips-unvectored-crash-'))
try {
let brain = await openKillMatrixBrain(dir, { logAuthority: 'adopt' })
// Baseline: real-vectored entities, durably flushed.
const baselineIds: string[] = []
for (let i = 0; i < 5; i++) {
const id = uid(`baseline-${i}`)
await brain.add({ id, data: `baseline entity ${i}`, type: NounType.Document, vector: vec(i) })
baselineIds.push(id)
}
await brain.flush()
// Crash window: at-ack writes that are never flushed before the crash.
const crashedIds: string[] = []
for (let i = 0; i < 4; i++) {
const id = uid(`crashed-${i}`)
await brain.add({ id, data: `crash-window entity ${i}`, type: NounType.Document, vector: vec(100 + i) })
crashedIds.push(id)
}
await abandonAsCrashed(brain)
// Reopen — logAuthority: 'adopt' replays the at-ack log for the crash window.
brain = await openKillMatrixBrain(dir, { logAuthority: 'adopt' })
for (const id of [...baselineIds, ...crashedIds]) {
expect(await brain.get(id), `entity ${id} must survive the crash`).not.toBeNull()
}
// Repair — must not disturb any entity, and must recount the ledger honestly.
const report = await brain.repairIndex()
expect(report.families.length).toBeGreaterThan(0)
for (const id of [...baselineIds, ...crashedIds]) {
expect(await brain.get(id), `entity ${id} must survive repair`).not.toBeNull()
}
const ledger = await brain.storage.getCanonicalCounts()
expect(ledger.suspect).toBe(false)
expect(ledger.vectors.all).toBe(baselineIds.length + crashedIds.length)
// Second life: a fresh real-vectored add must never trip a dimension
// mismatch against a vector-less phantom left in the index — the exact
// mechanism `clear-persistence.test.ts` and the biography lane hit.
const secondLifeId = uid('second-life')
await expect(
brain.add({ id: secondLifeId, data: 'second life entity', type: NounType.Document, vector: vec(200) })
).resolves.toBe(secondLifeId)
expect(await brain.get(secondLifeId)).not.toBeNull()
await brain.close()
} finally {
fs.rmSync(dir, { recursive: true, force: true })
}
})
})

View file

@ -382,9 +382,17 @@ describe.sequential('lifecycle — the working store', () => {
},
// Every noun this biography ever adds carries an explicit/computed
// vector (the harness never defers an embed), so the vectored-noun
// scalar tracks nouns.all exactly.
// scalar tracks nouns.all exactly EXCEPT for the VFS root counted
// in `vfsBaselineNouns`: the root is deliberately persisted with
// `vector: []` (the sanctioned "unvectored" shape — see
// VirtualFileSystem.doInitializeRoot()'s zero-norm-avoidance
// comment) so it never pays the WASM engine's cold-compile cost and
// never crosses an engine boundary as a false attractor. It is the
// ONE hidden-tier record `vfsBaselineNouns` represents (see
// biographyHarness's module header), so it is excluded here even
// though it counts toward `nouns.all`.
vectors: {
all: aliveEntities.length + model.vfsFileNouns + model.vfsBaselineNouns
all: aliveEntities.length + model.vfsFileNouns
},
suspect: false
})