Brainy 7.30.0 introduced a memory-derived synchronous cap on `find({ limit })`
to prevent OOM. The cap was sound in intent but ~4x too conservative in
calibration: assumed 100 KB per result while typical entity footprint is 7-10 KB
(384-dim float32 vector ≈ 1.5 KB + standard fields + metadata). On a 900 MB
free-memory box the cap derived to 9000 — breaking common safety-cap patterns
like `find({ type, where, limit: 10_000 })` that typically return 10-500
entities. Surfaced as a runtime regression with cascading 500s degrading
production dashboards.
Three concurrent fixes:
A. RECALIBRATE THE FORMULA
- src/utils/paramValidation.ts:175,196,212 — the three memory-derived priorities
(reservedQueryMemory / containerMemory / freeMemory) all divided by
100 * 1024 * 1024 (100 KB per result, ~10-15x over conservative). Replaced
with a new MAX_LIMIT_KB_PER_RESULT = 25 constant that matches observed
entity size.
- Result: 4 GB container cap goes 10_000 → 40_000; 2 GB cap goes 5_000 →
20_000; 900 MB free-memory cap goes 9_000 → ~36_000. 100k hard ceiling
unchanged. `maxQueryLimit` / `reservedQueryMemory` constructor overrides
unchanged in behavior.
B. TWO-TIER ENFORCEMENT (warn-then-throw)
- Below cap (limit <= maxLimit): silent pass, unchanged.
- Soft tier (maxLimit < limit <= 2 * maxLimit): NEW — one-time warning per
call site (dedup keyed on caller stack frame + limit value), query
proceeds. Pre-7.30.2 code that relied on the cap silently allowing typical
safety-cap limits keeps working; the warning teaches the recipe so consumers
can fix it intentionally.
- Hard tier (limit > 2 * maxLimit): throw with the same teaching message
format. Real OOM territory; the cap stops being a recommendation and becomes
a guardrail.
- The 2x soft margin absorbs typical safety-cap patterns (limit: 10_000
against a 9 K-cap box) without disabling OOM protection. Real OOM territory
on a JS in-memory brain is hundreds of thousands of results, not 10x the
safety cap.
C. IMPROVED ERROR / WARNING MESSAGE
- Same shape as the 7.30.1 enforcement-error messages: state the problem,
name the three escape valves (maxQueryLimit / reservedQueryMemory /
pagination), include caller location, link to docs.
- Extracted findCallerLocation() helper from brainy.ts to a new
src/utils/callerLocation.ts so both the subtype enforcement (7.30.1) and
the limit enforcement (7.30.2) share one implementation without circular
imports.
DOCS
- New docs/guides/find-limits.md (public: true) — full reference: why the cap
exists, the four memory sources the auto-config considers, the three escape
valves with when-to-use-which guidance, and an explicit "pagination is the
future-proof pattern" callout (8.0 may tighten the cap further; pagination
keeps working unchanged).
- docs/api/README.md find() entry gets a one-paragraph `limit` tip + pointer
to the new guide.
- RELEASES.md v7.30.2 entry.
TESTS
- New tests/integration/find-limits.test.ts (9 tests): below-cap silent pass;
soft-tier warns once per call site (dedup verified by exercising same vs.
different source lines via wrapper closures); soft-tier message format
(names all three escape valves + docs link); soft-tier message includes
caller location; hard-tier throws; hard-tier message format same as
soft-tier; consumer maxQueryLimit override raises the cap and shifts both
tiers accordingly; pre-7.30.2 regression scenario explicitly covered.
- tests/unit/utils/memoryLimits.test.ts — 4 tests updated for the recalibrated
cap values (hardcoded expected numbers bumped 4x to match new 25 KB/result
assumption).
- tests/unit/utils/paramValidation.test.ts — auto-limit test extended to cover
the three-tier semantics (below-cap pass / soft-tier silent / hard-tier
throw).
- Existing suites unchanged: subtype-and-facets 26/26, verb-subtype-and-
enforcement 30/30, strict-mode-self-test 13/13. Unit 1468/1468.
CORTEX COMPATIBILITY
- Zero Cortex changes required. Every change is JS-side: formula recalibration
runs in ValidationConfig.constructor(), two-tier enforcement runs in
validateFindParams(), both fire before any storage / index / Cortex call.
- The new guide notes that Brainy 8.0's Datomic-style Db.find() may tighten
per-call limits to keep snapshot semantics cheap; pagination remains the
pattern that's guaranteed to keep working.
REPO-WIDE CLEANUP
Brainy is the only Soulcraft project that is open source. This commit also
scrubs closed-source product names and product-specific class/field references
from every tracked file in the repo (src/, docs/, tests/, RELEASES.md,
CHANGELOG.md). Consumer-reported bugs, regression scenarios, and release
notes now refer to "a consumer", "a downstream application", "a production
deployment", or "an internal report" — never to the named product. Two
product-named test files renamed to neutral diagnostic names. CLAUDE.md gains
a project-level guard rule documenting the policy and an example list of the
identifiers that may not appear in tracked code.
Verification
- npx tsc --noEmit: clean
- npm test: 1468 / 1468 unit
- All four integration subtype + verb + strict + find-limits suites: 78/78
- npm run build: clean
- Closed-source product reference audit: clean
Brainy 7.30 shipped opt-in subtype enforcement; SDK 3.20.0 then registered
SDK_CORE_VOCABULARY on every consumer's brain (Event, Collection, Message,
Contract, Media, Document NounTypes). On 2026-06-08 Venue's /book flow went 500
because their brain.add({ type: NounType.Event, ... }) call sites lacked
subtype. An audit of Brainy's OWN source revealed 14 HIGH-risk internal write
paths that also omit subtype — any consumer running the same vocabulary would
have hit Brainy's infrastructure paths next. 7.30.1 closes both gaps before
8.0 makes strict mode the default.
Additive across the board. Zero behavior change for consumers not using strict
mode. Every change is JS-side — Cortex needs no work for 7.30.1.
NEW — brain.audit() diagnostic
- Read-only method walking storage.getNouns() / getVerbs() pagination
- Returns { entitiesWithoutSubtype: { type: count }, relationshipsWithoutSubtype,
total, scanned, recommendation }
- VFS infrastructure entities excluded by default (they bypass enforcement via
isVFSEntity marker); pass { includeVFS: true } to surface them
- The companion to migrateField (7.x) and fillSubtypes (8.0): tells consumers
exactly what would break under strict enforcement, deterministically
NEW — Improved enforcement error messages
- Caller's source location extracted from Error().stack so users see their own
call site, not a Brainy internal frame
- Specific guidance branches: registered vocabulary → "Pass one of: a, b, c";
brain-wide strict mode → mentions the except clause; otherwise → registration
recipe via brain.requireSubtype()
- Documentation link to the canonical migration recipe
- Same shape for noun and verb enforcement
NEW — CLI --subtype flag
- brainy add and brainy relate gain -s/--subtype <value>
- Defaults to 'cli-add' / 'cli-relate' so the CLI works against strict-mode
brains without the user needing to know the vocabulary in advance
INTERNAL — every Brainy write path now sets subtype
- VFS Contains edges (5 sites at lines 503/905/1694/1772/1886) → 'vfs-contains'
- VFS symlink entity → 'vfs-symlink' (NEW — distinct from 'vfs-file')
- VFS copy-file → preserves source subtype, falls back to 'vfs-file'
- VFS symlink also adopts the isVFSEntity infrastructure marker so it bypasses
enforcement in strict mode
- Aggregation materializer (Measurement entities) → 'materialized-aggregate'
- ImportCoordinator (3 sites): document → 'import-source'; entities →
options.defaultSubtype ?? 'imported'; placeholder → 'import-placeholder'
- SmartImportOrchestrator (4 entity sites + 2 batch relate sites): same
precedence (extractor → options.defaultSubtype → 'imported')
- EntityDeduplicator → candidate.subtype ?? 'imported'
- UniversalImportAPI → extractor → 'extracted' for both entities and relations
- NeuralImport → adds defaultSubtype to NeuralImportOptions; precedence same
- GoogleSheetsIntegration → request body 'subtype' ?? 'imported-from-sheets'
- ODataIntegration → request body 'Subtype' ?? 'imported-from-odata'
- MCP client message storage → 'mcp-message' (also fixes pre-existing missing
data field and missing type by aliasing from the prior text field)
Side-effect fix: storage.getNouns() paginated now surfaces subtype to top-level
- Single-noun getNoun() already did this in 7.30; the paginated path was missed
- Without this fix brain.audit() saw missing subtype on entities that actually
had one (caught by the strict-mode self-test before release)
NEW — tests/integration/strict-mode-self-test.test.ts (13 tests)
- Creates a brain under the exact SDK_CORE_VOCABULARY shape Venue hit + brain-
wide strict mode
- Exercises every internal Brainy path: VFS root + mkdir + writeFile + cp + mv
+ ln + symlink; aggregation engine; audit diagnostic with includeVFS toggle
- Validates error message UX: caller location, vocabulary guidance, brain-wide
strict mode guidance, off-vocabulary value reporting
Docs
- New "Strict mode in practice" section in docs/guides/subtypes-and-facets.md
covering the SDK_CORE_VOCABULARY pattern, 4-step migration recipe
(audit → migrateField → hand-fix → re-audit), the Brainy-internal label
reference table, and an 8.0 forward-look on fillSubtypes()
- docs/api/README.md: new audit() entry, strict-mode tips on add() and relate()
- RELEASES.md: full 7.30.1 entry
Cortex parity (forward-looking, not blocking 7.30.1)
- 6th open question added to .strategy/BRAINY-8.0-SUBTYPE-CONTRACT.md: native
fast path for audit() and fillSubtypes() via column-store null-subtype
bitmap for billion-scale brains
- Cortex should add a parity test mirroring strict-mode-self-test.test.ts
against their native paths to catch any latent bug where native writes
bypass JS validation
- Brainy-internal subtype labels become a documented part of the 8.0 contract
(useful for Cortex telemetry surfacing Brainy-managed infrastructure %)
Verification
- npx tsc --noEmit: clean
- npm test: 1468/1468 unit
- 7.29 noun integration suite: 26/26 (no regression)
- 7.30 verb subtype + enforcement integration suite: 30/30 (no regression)
- New strict-mode-self-test integration suite: 13/13
- npm run build: clean
- Closed-source product reference audit: clean
Addresses VE-SUBTYPE-MIGRATION (Venue's reported request) and ships internal
labels Venue did NOT ask for but that would have broken them next under their
own vocabulary registration.
Promotes `subtype?: string` to a top-level standard field on every entity,
alongside `type` / `confidence` / `weight`. Flat string, no hierarchy — the
consumer-chosen vocabulary for sub-classifying entities within a NounType
(Person → employee/customer, Document → invoice/contract, etc.).
Layer 1 — subtype field + rollup
- HNSWNounWithMetadata.subtype + STANDARD_ENTITY_FIELDS entry
- Entity / Result / AddParams / UpdateParams / FindParams threading
- add()/update() persist subtype on storageMetadata + entityForIndexing
- get()/find() route through the standard-field fast path
- subtypeCountsByType (Map<NounTypeIdx, Map<subtype, count>>) on
BaseStorage, mirrored after nounCountsByType with the same self-heal
rebuild and persisted to _system/subtype-statistics.json
- brain.counts.bySubtype(type, subtype?) — O(1) point + breakdown
- brain.counts.topSubtypes(type, n) — top-N by count
- brain.subtypesOf(type) — distinct subtypes seen
- find({ type, subtype }) and find({ subtype: ['a','b'] }) on the fast path
Layer 2 — trackField for other facets
- brain.trackField(name, { perType?, values? }) registers a field for
cardinality + per-NounType breakdown stats. Backed by the aggregation
engine (auto-defines __fieldCounts__<name>), backfill-on-define applies.
- brain.counts.byField(name, { type? }) returns value frequencies
- Optional vocabulary whitelist rejects off-vocabulary writes at add/update
Layer 3 — generic migrateField
- brain.migrateField({ from, to, readBoth?, batchSize?, onProgress? })
streams every entity, copies the value from one path to another, and
(unless readBoth) clears the source. Supports top-level standard fields,
metadata.X, and data.X paths. Idempotent — safe to re-run.
Docs
- New guide: docs/guides/subtypes-and-facets.md (Layer 1 + 2 + 3)
- README, DATA_MODEL, QUERY_OPERATORS, api/README, finite-type-system,
quick-start all treat subtype as a core primitive with anonymous example
vocabularies (employee/customer/invoice/milestone).
Tests
- 26 new integration tests covering write/read/update/delete round-trips,
counts rollup decrement + re-route on mutation, trackField + byField
with and without perType, vocabulary whitelist enforcement, and
migrateField for metadata.X → subtype and data.X → subtype paths
including readBoth deprecation-window semantics.
Unit suite: 1468/1468 passing. Type-check + build clean.
The JS fallback EntityIdMapper caps at u32::MAX to match the metadata
index's Roaring32 bitmap width. Before this guard, a brain past 4.29 B
entities would silently widen `nextId` into JS's safe-integer range,
truncating ints inside the bitmaps and zeroing query results — the
same silent under-count cortex 3.0's Piece 10 just closed on the
native path. Brainy 8.0 surfaces the overflow loudly instead.
When `nextId` would exceed `U32_ENTITY_ID_MAX`, `getOrAssign` throws
`EntityIdSpaceExceeded` with a message pointing at the cortex 3.0
binary mapper's `idSpace: 'u64'` mode as the migration path (mmap-
backed extendible-hash KV, persists the full u64 range losslessly).
The existing-UUID lookup path bypasses the guard — only fresh
allocations can overflow.
Surfaces via `@soulcraft/brainy/internals`:
- `EntityIdMapper` (already exported, behavior gated)
- `EntityIdSpaceExceeded` (new)
- `U32_ENTITY_ID_MAX` (new constant, = 0xFFFF_FFFF)
This is the lockstep half of cortex 3.0 / Piece 10 / Step 15. Cortex's
`NativeBinaryEntityIdMapperWrapper` ships the U64 binary mapper that
takes over above this ceiling; brainy ships the bright-line failure
that points consumers at it.
New test: tests/unit/utils/entity-id-mapper-u32-ceiling.test.ts (6
assertions covering the constant, the error shape, normal
allocations, the boundary case at exactly u32::MAX, the overflow
throw, and the existing-uuid lookup bypass).
createIndex() now consults the 'diskann' provider before 'hnsw':
1. config.index.type === 'diskann' → require the provider, throw if absent.
2. Auto-engage when ALL of:
- 'diskann' provider registered (cortex plugin loaded)
- storage.getBinaryBlobPath('_diskann/main') returns a local path
(cloud adapters return null → silently stay on HNSW)
- metadataIndex has a stable idMapper
3. config.index.type === 'hnsw' → force the historical in-memory engine.
4. Fall back to the cortex 'hnsw' provider, then brainy's TS HNSWIndex.
migrateToDiskAnn(options): builds the new index in parallel from the
current vector set, samples queries to verify recall ≥ recallTarget
against the old index, atomically swaps if recall passes. Throws and
leaves the old index in place otherwise.
migrateToHnsw(): always-available reverse path. Both preserve
canonical storage (entity JSON, metadata index) — only the search
engine changes.
The orchestration is generic — it works against any registered
'diskann' provider, not tied to any specific implementation. Brainy
without cortex falls back to HNSW; cloud-storage users transparently
stay on HNSW.
Adds the plugin-side surface for the new billion-scale index option:
- plugin.ts: DiskAnnProvider type alias (mirrors HnswProvider's shape
so cortex's DiskANN wrapper is a drop-in for the 'hnsw' factory).
- coreTypes.ts: HNSWConfig.type ('hnsw' | 'diskann') for explicit
engine selection, plus HNSWConfig.diskann for the build-time tuning
knobs (pqM, pqKsub, maxDegree, searchListSize, alpha, mmap
adjacency).
No algorithm code here — the implementation lives in the cortex
plugin. Brainy without cortex continues to use HNSW exactly as before.
Brainy now has a complete reference SQ4 (4-bit per dimension, 8× compression
vs float32) quantization layer in src/utils/vectorQuantization.ts: quantizeSQ4,
dequantizeSQ4, distanceSQ4Js, serializeSQ4, deserializeSQ4 — all byte-for-byte
identical to cortex's Rust quantize_sq4 / dequantize_sq4 / cosine_distance_sq4.
The pack format mirrors cortex exactly:
- Two nibbles per byte. High nibble = vector[2i], low nibble = vector[2i+1].
- Odd-dim vectors place the final value in the high nibble and pad the low
nibble with 0. Packed length = ceil(dim / 2).
- Zero-range vectors (all values identical) map every nibble to 8 (midpoint
of [0, 15]) — both reference impls produce 0x88 bytes.
The active-fn dispatch pattern matches SQ8: distanceSQ4Js is the pure-JS
reference; distanceSQ4 routes through a swappable activeSQ4Distance slot;
setSQ4DistanceImplementation(fn) swaps in cortex's SIMD Rust distance when the
distance:sq4 provider is registered. brainy.ts wires the provider in init,
same shape as the existing SQ8 wiring (~14 LOC). The dispatch path is the
single point of integration so HNSW search code never needs to know whether
it's running on JS or native.
Serialization adds a uint32 dim field after min/max (12-byte header total) —
SQ8 doesn't need it because the packed length equals dim, but SQ4's packed
length rounds up so dim must be recorded explicitly to disambiguate the
trailing pad nibble.
Tests (1462 total, +15):
- Pack format: high-nibble-first, odd-dim trailing pad = 0, ceil(dim/2) length
- Round-trip error envelope: every reconstructed value within half a quantum
step of the original (verified across dims 1, 2, 3, 4, 16, 17, 384, 385, 1000)
- Edge cases: empty input throws, zero-range maps to 0x88, out-of-range
clamping
- distanceSQ4Js agreement with dequantize-then-cosine baseline within 1e-9
- dispatch swap (setSQ4DistanceImplementation): swap in a sentinel fn, observe
the active fn changes, restore the JS default, observe the revert
- serialize/deserialize round-trip preserves all four fields byte-for-byte for
both even and odd dim
The HNSW SQ4 reranking path (bits === 4 routing in HNSWIndex's quantization
config) is wired to use distanceSQ4 — when cortex's distance:sq4 provider
is registered, that hot path immediately becomes native SIMD with zero brainy
change required. Cross-language byte-format parity tests run in the cortex
test suite (paired release brainy 7.28.0 + cortex 2.5.0).
BlobStorage.write() in `auto` compression mode now consults the new
`BlobWriteOptions.mimeType` and skips zstd for MIME types known to be
already heavily compressed — JPEG, PNG, WebP, MP4, WebM, MP3, ZIP, PDF,
Office formats, etc. Gzip/zstd over these formats wastes CPU for no
measurable byte savings; the payload entropy is already near maximal,
so the output is the same size or slightly larger plus the cost of
running the compressor.
The denylist `ALREADY_COMPRESSED_MIME_TYPES` is a conservative set of
well-known formats. False negatives (compressing something we should
have skipped) waste CPU; false positives (skipping something we could
have compressed) waste a few percent of bytes. The denylist favours
CPU-cycle safety because the formats listed here are the ones where
gzip/zstd is reliably a net loss.
The policy applies only to `auto` mode. Explicit `'zstd'` and `'none'`
are honoured because the caller is asserting the choice. The new
`isAlreadyCompressedMimeType()` is exported for consumers that want to
make the same decision before calling `write()`.
VFS / consumer wiring (pass mimeType from VFS through to BlobStorage)
is part of the 2.5.0 #27 storage unification work — when that lands,
every media upload through VFS will engage this policy automatically.
For now consumers opt-in by passing `mimeType` in BlobWriteOptions.
Tests (1447 total, +10):
- isAlreadyCompressedMimeType helper: canonical types, case-insensitive
+ parameter stripping, false-on-missing.
- write() auto-mode: image/jpeg + video/mp4 + application/zip all skip
compression (metadata.compression === 'none'); text/plain is allowed
through to zstd (either 'zstd' or 'none' depending on optional dep).
- Read decompresses transparently regardless of write-side decision.
- Explicit compression options bypass the policy.
Cortex registers a graph:compression provider exposing `encode`/`decode`
for HNSW connection lists (delta-varint, ~4× smaller edges than the
JSON-UUID-array shape brainy has persisted since the beginning). This
wires the consumer side without changing the saveHNSWData/getHNSWData
adapter signatures.
Architecture:
- NEW ConnectionsCodec (src/hnsw/connectionsCodec.ts) — translates between
UUID-keyed Map<level, Set<UUID>> and a single compact per-node buffer
via the provider + stable EntityIdMapper. Custom wire format:
[count: u8] [[level: u8] [len: u32 LE] [bytes...]]*. One blob per node
regardless of level count, so the storage I/O is identical to the
legacy path (one saveHNSWData call) plus a single saveBinaryBlob.
- HNSWIndex changes — a `connectionsCodec: ConnectionsCodec | null` field
with `setConnectionsCodec()` setter. THREE save sites (deferred flush,
immediate-mode entity persist, immediate-mode neighbor updates) now go
through a single `persistNodeConnections(nodeId, noun)` helper: when the
codec is wired AND the storage adapter exposes saveBinaryBlob, it
encodes the connections, stores them at `_hnsw_conn/<nodeId>`, and
records `connections: {}` in saveHNSWData as the marker. Otherwise the
legacy JSON-array path is taken. TWO load sites use a matching
`restoreNodeConnections` helper that tries loadBinaryBlob first and
falls back to the legacy hnswData.connections field on miss / decode
error. Format convergence is lazy: pre-2.4.0 nodes still load via the
legacy path, then write the compressed form on next dirty save.
- brainy.ts wireConnectionsCodec — runs alongside wireMmapVectorBackend
during init. Activates when (a) the graph:compression provider is
registered and (b) the metadata index exposes its idMapper. Unlike the
mmap-vector backend, this layer engages on EVERY brainy 7.25.0
adapter — the blob primitive itself is universal; only the codec
presence gates activation.
- Provider interface GraphCompressionProvider in plugin.ts — encode +
decode static signatures. Brainy depends on the interface; cortex's
registered { encode: encodeConnections, decode: decodeConnections }
satisfies it structurally.
Tests (1437 total, +4 vs prior tip):
- tests/unit/hnsw/connections-codec.test.ts — 4 unit tests with a JSON
mock provider: single-level round-trip, empty-Map round-trip (one-byte
buffer), multi-level fan-out round-trip, and silent-drop of UUIDs the
decoding idMapper no longer knows. The real cross-language byte
format is exercised when cortex 2.4.0 wires its delta-varint
encode/decode in.
This completes the brainy half of the 2.4.0 storage foundation: stable
ids (#23), mmap vectors (#24), graph link compression (#25), and
column-store interchange (#26). Coordinated release as brainy 7.26.0 +
cortex 2.4.0 follows.
Brainy's JS ColumnStore.flushBuffer now writes segment payloads + the
per-field DELETED bitmap via the binary-blob primitive (saveBinaryBlob) at
the cortex-shared key convention `_column_index/<field>/L<level>-NNNNNN`
(suffix-free; adapter appends its own). Byte-for-byte identical to what
cortex's NativeColumnStore writes — JS and native engines now read each
other's segments with no envelope re-encoding.
Closes the gap that the cortex 2.3.1 read-side fallback opened: cortex was
reading both formats but the JS engine was still WRITING the legacy
{_binary, base64} envelope, so a fresh JS write always required the cortex
fallback to kick in (and migrate lazily on first read). With this commit
JS writes natively in the unified format, and cortex's fallback only
fires on indexes persisted by older brainy releases.
Backward compat:
- ColumnStore.loadSegmentCursor tries the raw blob first, then falls back
to the legacy `{_binary, base64}` envelope at the `.cidx` object-path.
Indexes written by pre-2.4.0 brainy keep loading correctly.
- ColumnStore.init's DELETED-bitmap load has the same dual-format read.
- Adapters without the binary-blob primitive (custom adapters that didn't
follow the 7.25.0 surface) fall through to the legacy envelope writer
too, so writes still succeed there.
Tests (1433 total, +5 vs prior tip):
- tests/unit/indexes/columnStore/column-store-interchange.test.ts —
pins down the contract: (1) flush writes the raw blob, NO legacy
envelope; (2) DELETED bitmap likewise; (3) a legacy-format on-disk
segment loads correctly via the fallback; (4) legacy DELETED bitmap
ditto; (5) round-trip in the new format.
- All 101 existing ColumnStore tests pass — the new write path is
exercised by the existing lifecycle tests (MemoryStorage has the blob
primitive, so the new branch fires).
Cortex already registers the vectorStore:mmap provider (its Rust
NativeMmapVectorStore), but brainy has never consumed it — preloadVectors
and getVectorSafe still go straight to storage.getNounVector for every id,
even when an mmap layer is available. This wires the consumer end.
Architecture:
- NEW MmapVectorBackend (src/hnsw/mmapVectorBackend.ts) — bridges brainy's
UUID-keyed vector reads to a int-slot mmap file via the
vectorStore:mmap provider. Slots are addressed by the stable int id
from the post-2.4.0 #1 EntityIdMapper (the foundation this depends on).
Auto-grows the file (doubling) when a write lands beyond capacity, so
HNSWIndex never has to think about sizing. The class never touches
per-entity storage — it owns only the mmap layer.
- HNSWIndex changes — adds a vectorBackend field + a setVectorBackend
setter. The vector read paths (preloadVectors, getVectorSafe) try the
mmap layer first; on a storage fallback hit, they LAZILY write back into
the mmap slot. An upgraded install converges to the zero-copy fast path
under live traffic — no big-bang migration step. The legacy per-entity
path is preserved and still used when no backend is set.
- brainy.ts wiring — a new private wireMmapVectorBackend() runs once
during init, after plugin activation + metadataIndex setup. It activates
the backend only when (a) the vectorStore:mmap provider is registered,
(b) the storage adapter resolves a real local path via
getBinaryBlobPath(), and (c) the metadata index exposes its idMapper.
Cloud adapters return null on (b) and the backend is silently skipped;
HNSWIndex's behaviour is then identical to pre-2.4.0.
- Provider interfaces in plugin.ts — VectorStoreMmapProvider and
VectorStoreMmapInstance document the contract cortex's class fulfils
(the class IS the provider — static factory methods). Brainy depends on
the interfaces, not on cortex; the structural match is verified when
cortex 2.4.0 picks up this brainy release.
Tests (1428 total, +11 vs pre-2.4.0):
- tests/unit/hnsw/mmap-vector-backend.test.ts — 6 unit tests with an
in-memory mock provider. Covers round-trip, batch reads with interleaved
misses, slot stability (no re-slotting on overwrite), file growth without
data loss, idempotent open, and null returns for unwritten slots. The
real perf integration with cortex's NativeMmapVectorStore is exercised
when cortex 2.4.0 wires this in.
- tests/unit/utils/entity-id-mapper-stability.test.ts — moved here from
tests/regression/ (which is NOT in the unit-config include glob, so the
five #23 tests were not actually being run by npm test). The unit
config matches tests/unit/**/*.test.ts.
The 2.4.0 #2 follow-up will be the chunked-segment layout for remote
storage adapters (S3 / R2 / GCS) where a single growing file doesn't fit
immutable objects. For 2.4.0 release: local-FS only.
Previously metadataIndex.rebuild() called idMapper.clear() which reset nextId
to 1 and renumbered every UUID by re-insertion order. Any consumer that had
persisted int-keyed data against the old map was silently invalidated — and
2.4.0's vector mmap store (#20), graph link compression (#21), and column-store
JS↔native interchange all need persisted int indices that survive a rebuild.
Remove the unconditional clear() in rebuild(). The rebuild already re-iterates
every entity via idMapper.getOrAssign(uuid), which returns the existing int
unchanged for known UUIDs. Stale UUID→int entries for entities no longer in
storage persist as harmless memory overhead; a dedicated prune step can be
added if it ever matters.
clearAllIndexData() — the explicit nuclear recovery path — keeps its existing
idMapper.clear() call (renumbering is intentional there), and now logs a
prodLog.warn making it explicit that any persisted int-keyed data is invalidated
and must be rebuilt from canonical sources.
Strengthened the EntityIdMapper class JSDoc to document the stability guarantee
as a contract — append-only getOrAssign, monotonic nextId, remove() leaves
permanent holes, rebuild() never renumbers, only clear() does.
Added tests/regression/entity-id-mapper-stability.test.ts pinning down the
five-point contract: (1) single-rebuild stability; (2) many-rebuild stability;
(3) post-rebuild adds get fresh monotonic ints; (4) removes leave permanent
holes — new entities never recycle; (5) clearAllIndexData() explicitly
renumbers (the documented destructive path).
Foundation for 2.4.0 #2-#4. Full test suite (62 files, 1417 tests) green.
The native SQ8 distance hook renamed the original distanceSQ8 to distanceSQ8Js
and added a dispatcher in its place, but left the old function's doc block
orphaned above distanceSQ8Js (two consecutive comments, the first dangling).
Merge them into one accurate block on distanceSQ8Js with dash-notation params.
Native accelerators (cortex) register providers for metadataIndex, graphIndex,
hnsw, entityIdMapper, cache, columnStore, and aggregation. Until now the exact
method/property surface brainy calls on each was implicit — a provider could
drop a member brainy depends on and only fail at runtime when that path ran.
This defines and exports the provider contracts from the stable
@soulcraft/brainy/plugin entrypoint:
- MetadataIndexProvider, GraphIndexProvider, HnswProvider,
EntityIdMapperProvider, CacheProvider — each typed as exactly the surface
brainy calls (optional/feature-detected members like HNSW setPersistMode and
enableCOW are intentionally excluded).
- Re-exports ColumnStoreProvider and AggregationProvider (and the aggregate
types) from the same entrypoint so a plugin author can import the whole
provider surface from one place.
Brainy's own baseline classes now `implements` these contracts
(MetadataIndexManager, GraphAdjacencyIndex, HNSWIndex, EntityIdMapper,
UnifiedCache), so the interfaces can never silently diverge from what brainy
ships — and any provider that declares `implements` gets a compile error the
moment brainy starts requiring a new member.
The embeddings/embedBatch providers stay typed by the existing EmbeddingFunction
(no duplicate interface added). Type-only changes; no runtime behavior change.
Adds a swappable sort:topK seam for the find() result-ranking hot path. Brainy
ranks candidate results by relevance score then slices to offset+limit; for large
candidate sets that is an O(N log N) full sort even though only the page is kept.
New utils/resultRanking.ts exposes rankIndicesByScore (top-K index selection) with a
JS default (sortTopKIndicesJs) that is byte-identical in ordering to the previous
results.sort((a, b) => b.score - a.score) + slice: descending by score, stable ties
by original index (ES2019+ stable sort semantics). setSortTopKImplementation lets a
native provider (e.g. cortex's Rust partial-sort) replace it; the provider returns
indices into a scores array and only has to match the JS index ordering.
brainy.ts resolves the sort:topK provider in init() (same pattern as distance:sq8)
and wires both relevance-score sort sites in find() through rankIndicesByScore +
reorderByIndices. rankIndicesByScore validates a native provider's output (length,
range, no duplicates) and falls back to JS on any inconsistency, so a query never
returns wrong or duplicated results. No provider registered = unchanged JS behavior.
Tests: unit coverage of the dispatcher (JS ordering vs Array.sort across 200 random
trials incl. ties, provider routing, and fallback on bad/throwing providers) plus
find() integration proving ranking routes through a registered provider, that the
provider and JS fallback produce identical ids/scores on the same data, and that
pagination requests k = offset + limit with descending ordering.
Makes distanceSQ8 swappable via setSQ8DistanceImplementation and wires brainy.ts to
install a registered 'distance:sq8' provider (e.g. cortex's Rust SIMD), with the JS
distanceSQ8Js as the default fallback. The provider signature is byte-compatible with
the native cosineDistanceSq8, so HNSW SQ8 reranking transparently uses native distance
when cortex is loaded. Native==JS numeric equivalence is asserted by the cross-language
parity suite.
Replaces localeCompare / raw relational operators with compareCodePoints (UTF-8
byte order) in the remaining persisted or native-facing comparison sites:
- SSTable sourceId sort + zone-map range check + binary search (graph LSM)
- COW TreeObject + RefManager name sorts (reproducible content hashes and ref
listings across environments)
- getSortedIdsForFilter (numeric-aware; code-point for strings) so the JS sort
fallback matches the native column store exactly
Deterministic across OS/Node/ICU and byte-identical to the native engine. No
migration: COW serialize/deserialize preserve stored order, so existing trees
keep their hashes; only new writes adopt code-point order.
Introduce a first-class binary-blob storage primitive on the StorageAdapter
contract and implement it across all storage backends. This stores opaque byte
payloads verbatim instead of base64-in-JSON, eliminating the ~33% inflation and
full-materialization cost of the JSON envelope. It unblocks zero-copy,
mmap-able column-store segments and batch vector I/O at billion scale.
New methods (declared abstract on BaseStorageAdapter, the class that implements
StorageAdapter, and added to the StorageAdapter interface):
saveBinaryBlob(key, data) raw write, atomic on real filesystems
loadBinaryBlob(key) exact bytes, or null if absent
deleteBinaryBlob(key) idempotent (missing is ignored)
getBinaryBlobPath(key) real local fs path where one exists, else null
Shared key -> location convention across every adapter: the key's
"/"-separated segments nest under a `_blobs/` prefix and are suffixed with
`.bin`, e.g. "graph-lsm/source/sstable-123" ->
"<root>/_blobs/graph-lsm/source/sstable-123.bin". Blobs are not branch-scoped
(COW): they are immutable producer-managed segments.
Per-adapter behavior:
- FileSystemStorage: writes under <rootDir>/_blobs via tmp+rename; returns the
real on-disk path so native code can mmap it directly. Path convention matches
the existing MmapFileSystemStorage subclass byte-for-byte.
- S3CompatibleStorage / R2Storage / GcsStorage / AzureBlobStorage: put/get/delete
raw octet-stream objects; getBinaryBlobPath returns null (remote stores have no
local path).
- MemoryStorage: defensive-copied Map<string, Buffer>; null path; cleared on
clear().
- OPFSStorage: stores raw bytes in the OPFS tree; null path.
- HistoricalStorageAdapter: read-only — save/delete throw; load resolves the
blob from the historical commit tree; null path.
Tests: tests/unit/storage/binaryBlob.test.ts exercises save/load round-trip
(byte-identical, incl. non-UTF8 bytes), overwrite, delete-then-load, load-missing,
and getBinaryBlobPath behavior for all eight adapters. Cloud adapters run against
in-memory client fakes that drive the real adapter code; OPFS runs against an
in-memory FileSystem Access API mock; the historical adapter commits a blob into
a real COW tree. 59 new tests; full unit suite (1398 tests) green.
Replace localeCompare (default-locale, non-deterministic across environments —
unsafe for a persisted sorted index) with UTF-8 byte / code-point order via a
shared compareCodePoints() helper. String ordering is now deterministic and
byte-identical to the native column store / aggregation sort, so results are the
same with or without the native accelerator. Covers the aggregate orderBy sort
and all 5 column-store comparison sites (tail buffer, merge sort, binary search).
Adds compareCodePoints unit tests.
Add 'percentile' (with a 'p' fraction in [0,1]) and 'distinctCount' to the
aggregation engine. Both are exact, computed from a per-metric value multiset
(MetricState.valueCounts) maintained incrementally and delete-safe; percentile
uses numpy-linear interpolation. The multiset is JSON-serializable so results
survive persistence. 35 aggregation unit tests pass.
- groupBy now supports { field, unnest: true }: an entity contributes once per
distinct element of an array field (tag frequency / faceted counts). Duplicate
elements on one entity count once; an empty/missing array joins no group. The
incremental add/update/delete paths fan out across the unnested groups.
- extractEntities/extractConcepts batch-embed the unique candidate spans in one
embedBatch call instead of one embed() per candidate (N sequential model calls);
falls back to per-candidate embedding if the batch fails. No behavior change.
New report APIs:
- brain.queryAggregate(name, params) returns the clean AggregateResult[] shape
({ groupKey, metrics, count }) directly, instead of the search-Result wrapper that
find({ aggregate }) returns.
- HAVING: find({ aggregate, having }) and queryAggregate filter groups by computed
metric values (e.g. revenue > 1000), complementing where (which filters group keys).
Evaluated per group: O(groups), independent of entity count.
Fixes (all reproducible on Node; surfaced under Cortex+Bun by Memory):
- Aggregate backfill-on-define: defining an aggregate over a store that already holds
matching entities returned []. It now backfills from existing entities on first query
(storage-agnostic via getNouns), so it works under durable backends that reopen
pre-populated. groupBy:['noun'] resolves to the entity type; find({ aggregate }) rows
expose groupKey/metrics/count at the top level.
- Multi-hop find({ connected: { depth, via } }) honors depth and via at every hop
(was 1-hop only, with verb filtering applied to hop 1 only); BFS bounded by limit.
- extractEntities no longer bleeds a neighbour's type indicator across candidates; each
candidate is typed by its own span. Also fixes the "Dr." title pattern.
Adds real-embedding regression tests; full unit suite green.
A lightweight tag is skipped by 'git push --follow-tags', so v$VERSION stayed
local-only and 'gh release create' failed at the end of the release. Use an
annotated tag (-a) so it pushes with the release commit.
Three advanced-API correctness fixes, all reproducible on Node (not Bun-specific):
- Entity/concept extraction returned []. SmartExtractor combined agreeing signals
with a weighted sum compared against an absolute 0.60 gate, so a confident
low-weight signal lost selection to a mediocre high-weight one that then failed
the gate, dropping the whole result. Select and gate on a normalized weighted
average instead. The public `confidence` option now controls the threshold (was
a dead hardcoded 0.60), and the embedding-signal timeout is raised 100ms -> 2000ms
so the neural signal is not silently dropped on slower runtimes.
- Multi-hop find({ connected }) returned only the 1-hop neighbour. executeGraphSearch
ignored depth/via; it now delegates to the depth-aware neighbors() BFS.
- find({ aggregate }) hid groupKey/metrics/count under .metadata, so callers
expecting AggregateResult saw empty rows. Expose those fields at the top level.
Adds real-embedding regression tests in tests/integration/advanced-apis-regression.test.ts.
Per the Cortex team's audit, the BR-CX-INTERFACE-GAP boot crash was a
build/install artifact (stale node_modules, lockfile drift, Docker cache,
bundler quirks), not a plugin-bundles-brainy version-skew issue. Cortex's
MmapFileSystemStorage extends FileSystemStorage and inherits all 6
multi-process helpers via the prototype chain at runtime; the dynamic
ESM import to @soulcraft/brainy is preserved in cortex's dist.
- Rewrote the hasStorageMethod() doc comment to credit the real cause.
- Updated the init-time warning to point operators at the actual fix:
clean install or container rebuild to refresh node_modules.
- New docs/concepts/storage-adapters.md documenting the inheritance
contract: extend FileSystemStorage to inherit the multi-process
helpers, override supportsMultiProcessLocking() -> true to activate.
Includes a minimum checklist for adapter authors.
- New docs/architecture/multiprocess-storage-mixin.md as a future-
direction note: extracting the 7 methods into a
MultiProcessSafeStorage interface/mixin is the architecturally clean
next step, deferred to v8 or until a second multi-process capability
lands.
No behavior changes. Ships with the next release.
Two production correctness defects fixed together so consumers can land
one upgrade.
BR-FIND-WHERE-ZERO — find() returned [] and stats() reported 0 entities
for any workspace whose data was written after the 7.20.0 column-store
refactor. Root cause: getStats() and the getIds() fallback still read
from the deleted sparse-index path. Separately, BaseStorage.getNounType()
was hardcoded to return 'thing', poisoning type-statistics.json with
every noun attributed to that bucket.
- MetadataIndex.getStats() reads from ColumnStore + idMapper.
- MetadataIndex.getIdsFromChunks() throws BrainyError(FIELD_NOT_INDEXED)
when neither store has the field. getIdsForFilter() catches per
clause, logs once, returns [].
- BaseStorage.nounTypeByIdCache populated in saveNounMetadata_internal
and consumed in saveNoun_internal. flushCounts() now persists the
Uint32Array counters too, so readers see fresh per-type counts.
- Self-heal at init: loadTypeStatistics() auto-rebuilds when the
poisoned signature is detected. rebuildTypeCounts() is now public for
use from `brainy inspect repair`.
- Dead state removed: dirtyChunks, dirtySparseIndices,
flushDirtyMetadata().
BR-DEFENSIVE-INTERFACE — 7.21.0 called supportsMultiProcessLocking()
unconditionally, crashing on older Cortex storage adapters that predate
the method. New hasStorageMethod(name) helper gates every new-method
call site. Older adapter triggers a one-line warning at init pointing
at the recommended plugin version.
Tests:
- new tests/integration/find-where-zero.test.ts (7 cases)
- new tests/integration/cortex-compat.test.ts (5 cases)
- multi-process-safety.test.ts updated to enforce correct counts
- 1329/1329 unit + 24/24 new integration tests passing
Filesystem storage now enforces single-writer, many-reader semantics.
A second writer on the same data directory throws at init time with the
holder's PID, hostname, and heartbeat — replacing the previous silent
stale-reads failure mode.
- New: `Brainy.openReadOnly()` — coexists with a live writer, every
mutation throws clearly.
- New: writer lock at `<rootDir>/locks/_writer.lock` with 10s heartbeat
and stale-detection (PID liveness + heartbeat freshness).
- New: cross-process flush-request RPC (filesystem-based, no signals)
so inspectors can force fresh state on demand.
- New: `brain.stats()`, `brain.explain(findParams)`, `brain.health()`
for operator-facing introspection.
- New: `brainy inspect` CLI with 13 subcommands (stats, find, get,
relations, explain, health, sample, fields, dump, watch, backup,
repair, diff), all read-only by default.
- Same-PID re-opens allowed with a warning (preserves test "simulate
restart" patterns).
- Storage instances passed directly via `storage: new MemoryStorage()`
are now honoured instead of silently falling through to the
filesystem auto-detect path.
Brainy + Cortex compose under this model — the lock covers both because
they share `rootDir`, Cortex segments are immutable mmap files, and
MANIFEST updates use atomic-rename.
Column store handles all writes and primary queries. Sparse index write
methods are unreachable:
Deleted: addToChunkedIndex (~103 lines), removeFromChunkedIndex (~37 lines),
splitChunkDeferred (~82 lines), getValueChunkFilename + makeSafeFilename
(~20 lines). Total: ~240 lines of dead write-path code.
Sparse index read methods kept as migration fallback for existing workspaces.
Will be deleted once lazy migration (buildFromStorage) ships.
Adds a per-field sorted column store (Lucene doc values + roaring bitmap
architecture) that replaces the MetadataIndex sparse index internals for
both filtering and sorting. One system for all field types with exact
precision — no bucketing, no per-entity storage reads.
Column store: binary .cidx segment format, in-memory tail buffers,
LSM-style compaction, k-way merge sort, multi-value support (__words__).
All queries (filter, range, sort, filtered sort) route through the column
store when data is available, falling back to sparse index otherwise.
Key unlocks:
- find({ orderBy: 'createdAt' }) works WITHOUT a filter (previously threw)
- find({ orderBy: 'metadata.price' }) works for custom numeric fields
- Exact timestamp precision (no 1-minute bucketing)
- O(K log S) sort independent of total entity count
New files: src/indexes/columnStore/ (types, format, tail buffer, cursor,
manifest, coordinator — ~700 lines). 101 new unit tests covering binary
format round-trips, CRC validation, sort, filter, range, deletion,
multi-segment merge, persistence, and multi-value (words) fields.
Deleted: metadataIndex-automatic-bucketing.test.ts (bucketing behavior
eliminated by exact-precision column store). Sparse index write path
removed from addToIndex/removeFromIndex. Sparse index legacy code still
present as dead code pending cleanup in next commit.
Preventive cleanup of sites that previously used dual-lookup patterns
(metadata[field] ?? entity[field]) or hardcoded timestamp if-chains
to read fields off HNSWNounWithMetadata. These worked today but were
fragile — the same failure mode that caused the orderBy sort bug
would have recurred the next time someone reordered or dropped a
lookup.
- AggregationIndex.computeGroupKey + getNumericField now route all
dimension and metric field lookups through resolveEntityField.
- improvedNeuralAPI._getItemsByTimeWindow + _clusterItemsByField
use resolveEntityField as the primary path, with item.data as a
legacy producer fallback. The type/nounType legacy branch is
preserved as-is since it handles storage format compatibility
rather than shape contract drift.
No behavior change for correct inputs. All aggregation and orderby
regression tests pass unchanged.
Cortex and other first-party plugins need a shared contract for reading
fields off HNSWNounWithMetadata. Exporting from /internals keeps a single
source of truth and prevents the shape contract from drifting between
packages.
