# Changelog
- RvLite CLI persists literal vector IDs such as __proto__ and requires own records for get/delete, preventing silent data loss and phantom inherited IDs; automatic allocation skips occupied caller IDs and refuses exhausted counters.
All notable changes to RuVector will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [@ruvector/core 0.2.0] - Unreleased
### Changed (breaking)
- `new VectorDb({ dimensions })` and `VectorDb.withDimensions(n)` without
`storagePath` now create an in-memory database private to that instance
(issue #1063). They used to open a shared `./ruvector.db` in the working
directory, so every such instance shared one store and its stored dimension
silently overrode the requested one: a 4-dim index created after a 256-dim one
accepted construction, then failed every insert with `Dimension mismatch`.
**Upgrading:** to keep an existing implicit store, pass
`storagePath: './ruvector.db'`. When `storagePath` is omitted and that file
exists, the binding prints a one-time warning to stderr.
- Opening an existing store whose stored `dimensions` differ from the requested
ones now throws in the constructor, naming the path and both values. An
explicitly passed `distanceMetric` is checked the same way; an omitted one
keeps the stored metric, so reopening a non-Cosine store without restating the
metric still works.
- Needs native `ruvector-core-*` platform packages built from this change;
publish them at the same version and bump the `optionalDependencies` pins.
Published as 0.2.0 rather than 0.1.x so caret (`^0.1.x`) dependents do not
pick up the new default silently.
## [ruvector-core] - Unreleased
### Added
- `VectorDB::open_checked(options, ConfigCheck)` and `ConfigCheck`
(`ALL`, `DIMENSIONS`, `NONE`): open or create a database, failing with
`InvalidParameter` when an existing store's `dimensions` and/or
`distance_metric` differ from the requested ones instead of adopting them. A
legacy file holding vectors but no stored config is checked against a stored
vector's length before the requested config is written into it.
### Changed
- `VectorDB::new` keeps its documented behaviour (the stored configuration is
adopted) and now logs a `tracing` warning when the stored `dimensions` or
`distance_metric` differ from the requested ones.
## [Unreleased]
### Fixed (Python SDK)
- Python LangChain adapter: duplicate caller IDs in one batch are rejected before inserting vectors, preventing rows that cannot be deleted through the adapter.
- Python Collection persistence binds new sidecars to their index bytes with a SHA-256 digest, rejecting mixed file pairs after an interrupted index replacement. Legacy sidecars remain readable without this integrity guarantee.
- Python MCP tools evict mutated collection state after a failed save so rejected writes cannot leak into reads or later retries.
- Python MCP: refuse collection creation when a native index already exists without its sidecar, preserving partially restored data.
- Python MCP: discover persisted collections from their metadata sidecars, including empty RaBitQ collections, and omit incomplete native-index files.
- Python LlamaIndex searches apply node, document, and metadata restrictions before selecting the nearest top-k results. Empty document restrictions now return no results.
- Python LlamaIndex IS_EMPTY filters now include absent and null metadata fields, matching the framework contract alongside existing empty-string and empty-list handling.
- Python Agentforce routes: enforce the configured per-collection vector cap before upserting, consistently with MCP writes.
- Python Salesforce ingestion: reject incomplete pages without cursors and repeated pagination cursors rather than returning partial results or looping.
- Python RaBitQ identity integrity also ignores missing deletions and preserves IDs, metadata, and subsequent insertion identities through vacuum, keeping adapter document maps valid.
- Python RaBitQ export and persistence retain caller-supplied external IDs instead of substituting internal row positions, preserving associated Collection metadata through save/load.
- Python CLI: empty insert-batch inputs now report zero inserted rows successfully instead of indexing an empty ID list.
- Python graph queries preserve unknown null predicates through comparisons, NOT, AND, OR and XOR, preventing missing properties from being admitted by negation; integer equality and ordering retain precision beyond the float mantissa.
- Validate Python GNN checkpoint declarations, tensor layout and finite parameters before loading while preserving valid serialized learned weights.
### Fixed
- RvLite: acquire browser writer leases in a committed atomic IndexedDB transaction and prevent expired holders from overwriting or releasing successor leases.
- RvLite CLI refuses corrupt or unreadable existing databases instead of reporting empty data or overwriting recoverable bytes.
- Memory-backed Rust batches validate all vector dimensions before storing any
- `rvf_store_open` refuses a malformed or dimensionless `.rvf` buffer instead of skipping the bad segment and allocating a dimension-1 store.
- Count undirected edge weights once when scoring modularity in the npm
graph helpers and returned minimum-cut partitions. Single-edge and
disconnected-component scores now use the correct total edge weight.
- npm rvf WASM queries use the metric stored on the database. Passing zero to `rvf_store_query` forced L2, so cosine and inner-product stores ranked with the wrong distance.
- npm rvf rejects a batch whose rows do not match the store dimension before flattening, so a short vector cannot pad into the next row.
- npm rvf keeps a quarantined id-map sidecar refused on the next open until a valid mapping is restored. A fresh store with no quarantine still opens.
## [ruvector 0.3.23] - 2026-11-10
### Fixed
- `LearningEngine.import()` accepts state without the `neural` field again.
0.3.22 typed its parameter as the full `export()` return type, so TypeScript
code passing state exported by, or typed against, 0.3.21 stopped compiling.
`neural` is now optional (`LearningEngineImportState`); runtime behaviour is
unchanged. A type-level test compiles a 0.3.21 `export()` shape.
- `@ruvector/edge-full` source: `modules.onnx.size` and `totalSize.withOnnx`
report the current ONNX wasm size (10.2MB, 11.5MB with ONNX) instead of
7.1MB / 8.4MB (not republished).
## [ruvector 0.3.22] - 2026-11-10
### Added
- `neural-q` learning algorithm: a small MLP Q-function (hashed state features,
64-32-|actions|) trained with AdamW on the experience buffer, so routing
generalises to states the tabular Q-table has never seen. Available in
`LearningEngine`, `ruvector hooks learning-config -a neural-q` and the
`hooks_learning_config` MCP tool. It runs on the optional peer `ruv-fann`
(`npm i ruv-fann@0.1.0`); without it `neural-q` behaves exactly like
`q-learning`. The default install is unchanged. Held-out benchmark
(`scripts/neural-q-bench.js`, 9 agents, 5 seeds): unseen-state routing
accuracy 0.924 vs 0.204 (route keys) and 0.597 vs 0.203 (free-text states)
for neural-q vs q-learning; about 7 us per decision vs about 1 us.
## [rvf-runtime 0.5.5] - 2026-11-10
### Added
- `RvfStore::should_compact`, `should_compact_with` and
`compact_if_needed` wire the spec 10 §7 compaction policy (dead space,
segment count, interval, emergency ratio) into the store. The policy and
`compaction::{CompactionThresholds, CompactionDecision, evaluate_triggers}`
were implemented and tested but unreachable (#755, from #753).
## [ruvector-sona 0.2.5, ruvector-graph 2.3.5] - 2026-11-10
### Fixed
- `MicroLoRA::set_weights` rejects NaN/Inf in either projection before
changing any state, so corrupt external weights cannot poison an adapter
(#1472, from #375).
- `GraphDB::delete_edge` and `delete_edges_batch` persist the deletion
before touching memory and indexes, so a storage error no longer leaves
memory and disk disagreeing (#1472, from #375).
### Documentation
- ruvector-graph README quickstart rewritten against the real `GraphDB` /
`NodeBuilder` / `EdgeBuilder` API (it used types that do not exist).
## [ruvector 0.3.21, ruvector-cli 2.3.1] - 2026-11-10
### Fixed
- `rvlite_sql`, `rvlite_cypher` and `rvlite_sparql` MCP tools work. They
constructed a nonexistent `rvlite.Database` (the class is `RvLite`), did
not await the async query methods, and stripped `( ) { } ' "` from queries
with a shell sanitizer although queries never reach a shell. `db_path`
stays confined to the working directory and is loaded, never written
back (#1463).
- Both MCP servers (`ruvector mcp start` and `ruvector-mcp`) answer
`resources/templates/list` with an empty list instead of -32601, which
clients such as Continue treat as fatal during setup (#1464, from #217).
## [ruvector 0.3.20] - 2026-11-09
### Fixed
- Attention wrappers call `@ruvector/attention` with its real argument
shapes (GraphRoPe, DualSpace, async flash/hyperbolic, parallel/batch
compute, InfoNCE, hard-negative mining, Adam, benchmark), and
`ruvector attention compute -t flash|linear` works (#1375).
- `hooks init` merges with other tools' hooks instead of replacing them,
hook commands read Claude Code's stdin JSON (`--stdin`), timeouts are in
seconds, and init no longer enables `claude-flow` or adds agentic-flow
alpha prompt hooks (#1222).
- Session hooks run on `SessionEnd` instead of after every reply,
`compressedPatterns` survive later saves, and `hooks compress` /
AUTO_COMPRESS store the Q-table (#1223).
- `hooks graph-cluster` / `graph-mincut` name nodes by relative path, so
same-named files in different directories stay distinct, and follow
`../x`, subpath and directory-index imports (#1217).
## [@ruvector/core 0.1.38 (platforms 0.1.35), ruvector 0.3.19] - 2026-11-09
### Fixed
- Re-inserting an existing id is an upsert with no HNSW ghost of the old
vector, and NaN/Inf vectors are rejected on insert, insertBatch and search
(#1443, issues #1310 / #1313). Core binaries built from b59a81a06.
- HNSW returns complete results on very small graphs and skips deleted
nodes (#1299). Shipped as core platform packages 0.1.35; 0.1.37 still
pinned the 0.1.34 binaries, built before the fix.
- linux-arm64-gnu core binary is baseline aarch64 again (#1384): 0.1.34
carried 2,524 LDAPR (`+rcpc`) instructions and died with SIGILL on
Cortex-A72 (Pi 4) and Jetson AGX Xavier. The 0.1.35 binary has none.
- `ruvector hooks build-agents` honours an absolute `--output` (#1294) and
`hooks rag-context --rerank` actually reranks (#1319).
- `@ruvector/rvf-mcp-server` 0.2.5 and `ruvector-extensions` 0.1.6 require
express ^4.22.3 (#1281).
## [ruvector 0.3.14] - 2026-11-06
### Security
- `decompile_url` private-address guard also checks IPv4 addresses embedded
in IPv6 forms (#1359).
- DiskANN build/search refuse symlinked index files (#1381).
### Changed
- Fresh installs resolve `@ruvector/rvf` 0.4.6 (filtered queries exact on
weak graphs, `forceExact`), `@ruvector/attention` 2.3.0 (native faster than
a JS loop, `computeFlat`, complete typings) and `@ruvector/core` 0.1.36
(`version()` reports the package version) through the existing ranges.
## [ruvector 0.3.13] - 2026-11-05
### Added
- Hook memory recall (`IntelligenceEngine.recall`, MCP `hooks_recall` /
`hooks_suggest_context`) uses a RaBitQ candidate index from
`@ruvector/rabitq-wasm` once a store holds `RUVECTOR_RABITQ_THRESHOLD`
memories (default 5000; `0` disables). The index proposes 10 x k candidates
and they are re-scored with exact cosine, so scores are exact and top-10
recall matched brute force (1.000) on 20k x 384 at ~2.7 ms vs ~38 ms per
query. The index is built on the second eligible recall in a process, so
one-shot hook commands never pay the ~0.4 s build. Post-build
inserts/deletes apply as a delta without a rebuild. Without the package,
recall is the unchanged exact scan. `@ruvector/rabitq-wasm` moves from
devDependencies to optionalDependencies (282 KB).
- `ruvector diskann build
-i vectors.json` / `ruvector diskann search
-v ...` and MCP tools `diskann_build` / `diskann_search` (paths
confined to the working directory) expose `@ruvector/diskann`, previously
library-only. It stays an opt-in extra: `npx ruvector add diskann`.
- `ruvector skills list|install` ships the `ruvector-integrations` agent
skill and installs it into `.claude/skills` (Claude Code) and/or
`.agents/skills` (Codex). `skills/` is now in the published files.
### Fixed
- Keep generated memory-storage IDs from overwriting live caller-selected rows, reserve explicit IDs in mixed batches, and refuse generated-counter wraparound while preserving intentional explicit upserts.
- Preserve current HNSW caller IDs across replacements and deletion, widen live top-k retrieval when obsolete graph rows crowd candidates, and repair stale reverse mappings when restoring older snapshots.
- Every open-ended `>=` peer range is now bounded. `@ruvector/ruvllm >=2.0.0`
resolved to 4.0.1, a different product published under the same name
without the runtime that `ruvector llm` / `ruvector sona` load (`ruvllm.RuvLLM is
not a constructor`); it is now `^4.1.0`, the restored runtime, verified
with `llm info|models|embed` and `sona status`. Other peers take their optionalDependencies spec
or a caret on their previous minimum.
- Security: `@metaharness/darwin` moves from `^0.9.1` to `^0.13.0` (and the
`metaharness` meta-package from `^0.4.7` to `^0.4.20`, which itself requires
darwin `^0.13.0`) for the high-severity sandbox bypass fixed in darwin 0.13.0
(GHSA-mjwg-5fm3-fx2p). The 0.x caret could never reach 0.13. `ruvector harness
darwin --execute` and `runMetaHarnessDarwin` call `evolve(config)`, whose
signature is unchanged; verified with a real evolution run against 0.13.0.
- Security: `@ruvector/attention` moves from `^0.1.3` (resolved 0.1.32) to
`^2.2.3`, which contains the H1 fix: on 0.1.x a malformed input panicked in
the NAPI binding and aborted the whole Node process (SIGABRT); 2.2.3 throws a
catchable error. The JS surface is identical (same 39 exports, byte-identical
wrapper); a differential run of every ruvector attention wrapper gave identical
outputs (max |diff| 1.8e-3 on hyperbolic, float drift). `@ruvector/core`,
`gnn` and `sona` already resolve to their latest. Guarded by
`test/attention-panic.js`.
- `harness status` no longer imports `@metaharness/darwin` to report it (darwin
0.13 eagerly loads `@babel/parser`, ~35 ms, which pushed status past the
ADR-256 startup budget: +135 ms vs 120 ms on CI). Status reports darwin as
installed, with its real installed version, without loading it; `harness
doctor` still imports every package. Delta back to ~+40 ms.
- Security: `@ruvector/sona` floor raised from `^0.1.4` to `^0.1.11`
(GHSA-c27m-v4wg-f9qj, affects < 0.1.9; 0.1.11 adds catch_unwind on every sync
NAPI entry point and a bounded trajectory store) so existing lockfiles move.
## [ruvector 0.3.12] - 2026-11-05
### Fixed
- `hooks graph-mincut` and `hooks graph-cluster` (and the MCP tools
`hooks_graph_mincut` / `hooks_graph_cluster` that shell out to them) failed
on every input: the CLI passed `(nodes, edges)` with `{source,target}` edges
to functions that take one `Graph` built from `{from,to}` edges. They now
build the graph correctly, recognise `require('./x')` and extension-ful
imports, and print clusters as JSON (`assignments`, `groups`, `modularity`)
instead of empty objects from serialized `Map`s.
- `graph-mincut` now uses `minCutFast`, which runs on the optional
`@ruvector/mincut-wasm` when installed and falls back to the exact TS
Stoer-Wagner otherwise; the output reports which `engine` ran.
- `louvainCommunities` could loop forever on any graph with a bridge edge: it
scored only the gain of joining a community, not the loss of leaving one, so
a bridge node oscillated. It now scores moves with the node removed from its
community, moves only on strict improvement, and normalizes by the
undirected edge weight like `calculateModularity`.
- `spectralClustering` used the largest eigenvectors of the normalized
Laplacian (power iteration) instead of the smallest, and seeded both power
iteration and k-means randomly. It now iterates on `2I - L` and seeds
deterministically, so it finds the real clusters and is reproducible.
## [@ruvector/router 0.1.32] - 2026-09-30
### Fixed
- `maxElements` was typed and documented as "Maximum number of elements" but
was never read: a database created with `maxElements: 5` accepted 12 inserts
(#1099). It is now documented as what it is, a **capacity hint that is not a
hard limit**, and a new opt-in `enforceMaxElements: true` (Rust:
`VectorDbBuilder::enforce_max_elements`) makes it a real bound. An insert
that would exceed the bound fails with `VectorDbError::CapacityExceeded`
(JS: an `Error` whose message starts `ERR_CAPACITY_EXCEEDED:`), before
anything is written; a batch is rejected as a whole. Replacing an existing
id never counts against the bound and deleting frees a slot. The check is
serialized so concurrent inserts cannot both pass it.
- Enforcement is opt-in rather than default because the default `maxElements`
is 1,000,000: enforcing it unconditionally would start rejecting writes to
any existing store past that size whose owner never asked for a limit.
`enforceMaxElements` without `maxElements` is rejected as a configuration
error.
## [@ruvector/router 0.1.31] - 2026-09-21
### Fixed
- Ship the issue #430 HNSW fixes (result-heap eviction order, insert beam width,
distance-based pruning, `k > efSearch`) that landed on `main` in May but never
reached npm: published 0.1.30 scored recall@10 = 3.6% on 5,000 real MiniLM
embeddings, returning neighbours from the last-inserted region regardless of
the query.
### Changed
- Neighbour selection now uses the HNSW paper's diversity heuristic (Algorithm 4,
with kept-pruned fill) for both new-node edges and overflow pruning, instead of
"the m closest". On clustered data the closest-m rule left almost no
inter-cluster edges, so single-entry-point search stalled inside a cluster
(recall@10 0.89 at defaults with clusters inserted contiguously). Measured on
5,000 real embeddings: recall@10 1.000 in every configuration tried (from
0.96–0.99 at defaults and 0.88–0.92 at m=16), with lower query latency
(p95 0.58 ms vs 0.85 ms at defaults). Inserts are 25–30% slower: the
heuristic computes candidate-to-kept distances at selection and at each
overflow prune, a per-insert cost bounded by `ef_construction × m` that does
not grow with corpus size.
- New regression test pins recall@10 ≥ 0.90 against brute force on clustered
data in both contiguous and shuffled insertion order, and fails if results
are drawn from the insertion tail — the failure mode uniform-random-vector
tests cannot see.
## [ruvector npm 0.3.2] - 2026-09-21
### Fixed
- `DbOptions` documented `hnsw: { m, efConstruction, efSearch }` and `path`,
but the `VectorDB` wrapper only read `hnswConfig` and `storagePath`, so a
caller following the published types had their HNSW settings silently
dropped and got the defaults. Both spellings are honoured; a partial `hnsw`
merges over the defaults; the canonical name wins if both are passed. The
resolution is a pure exported `resolveVectorDbOptions()` with a workflow test.
- `DbOptions` now documents the canonical native names (`dimensions`,
`hnswConfig`, `storagePath`) with the old spellings marked as aliases; the
never-implemented `autoPersist` field is gone.
- `ruvector route info` printed `v0.1.28` for every installed router version
(a hardcoded fallback); it now reads the installed package's version.
## [ruvector npm 0.2.37] - 2026-07-27
### Fixed
- Restore the release CI's native backend installation so functional,
learning, performance, and recall gates exercise HNSW instead of the
in-memory fallback.
- Preserve the Rust core's default HNSW behavior when JavaScript callers omit
`hnswConfig`; the wrapper previously selected the exact FlatIndex path.
- Normalize the package binary metadata and its integration assertion.
- Remove the npm workspace's critical `tar` advisory with a patched override.
- Correct the Shor factoring milestone table to distinguish a failed
35-factor experiment from the separate Schnorr/QAOA hybrid result.
## [ruvector npm 0.2.36] - 2026-07-26
### Fixed
- Start the stdio MCP transport when launched through `ruvector mcp start`
([#715](https://github.com/ruvnet/ruvector/issues/715)).
- Keep ONNX model-loader diagnostics on stderr so they cannot corrupt MCP
JSON-RPC frames ([#710](https://github.com/ruvnet/ruvector/issues/710)).
### Security
- Update `anyhow` to 1.0.104 and `memmap2` to 0.9.11, removing the
corresponding RustSec soundness exceptions.
- Update the MCP SDK to 1.29.x, clearing its shared-transport data-leak and
ReDoS advisories, and remove the unused `js-beautify` dependency.
## [hailo-backend] - 2026-05-03
Branch-only entry; not yet released to a versioned tag. 38 iters /
45 commits arc covering iter 133-171 to land NPU-accelerated
embedding inference on Pi 5 + AI HAT+ end-to-end.
### Added
- **NPU acceleration via Hailo-8** (ADR-176, iters 158-163). New
modules `hef_pipeline.rs`, `host_embeddings.rs`, `hef_embedder.rs`
in `ruvector-hailo`. `HailoEmbedder` auto-detects `model.hef` +
safetensors trio and routes through the NPU; falls back to
candle-on-CPU when HEF is absent.
- **First-known all-MiniLM-L6-v2 HEF for Hailo-8** — published as
GitHub Release `hailo-encoder-v0.1.0-iter156b` (15.7 MB, sha256
`cdbc89...`). Compiled iter 156b after working around four
distinct Hailo Dataflow Compiler v3.33 SDK bugs from user-space
(KeyError, AccelerasValueError, ElementwiseAddDirectOp Keras
serialize via `acceleras` Layer monkey-patch, RGB-align via
single-input encoder form).
- **`deploy/download-encoder-hef.sh`** — sha256-pinned downloader
for the HEF artifact, matches the iter-134
`download-cpu-fallback-model.sh` pattern.
- **`deploy/compile-encoder-hef.py`** — Python SDK driver for the
HEF compile pipeline, replaces the iter-131 CLI invocations that
hit the `hailo` CLI's `-y` auto-accept-recommendation bug.
- **`deploy/export-minilm-encoder-onnx.py`** — torch.onnx.export
helper that strips the BERT embedding lookup so the encoder
block compiles cleanly on Hailo-8.
- **iter-167 ranking-aware startup self-test** — worker now
embeds three reference phrases at boot and checks
`sim(close) > sim(far)`; refuses to serve if the encoder is
producing nonsense vectors.
- **Pi 4 / Pi 5-without-HAT deploy** as a first-class target
(ADR-177). cpu-fallback is fully hardware-agnostic across
aarch64.
- **iter-147 cpu-fallback embedder pool** —
`RUVECTOR_CPU_FALLBACK_POOL_SIZE=N` runs N parallel BertModel
instances behind try-lock dispatch. Measured 1.75× throughput
on x86 release, 4× on Pi 5 expected.
- **iter-143 fingerprint integrity** — `compute_fingerprint` now
hashes the safetensors+tokenizer+config trio so cpu-fallback
workers get integrity-checked by the cluster.
- **iter-141 cross-build with `--with-worker`** — produces an
aarch64 cpu-fallback worker binary in one command on x86.
### Performance
Measured on cognitum-v0 (Pi 5 + AI HAT+) at concurrency=4 via
cluster-bench:
| Path | Throughput | p50 latency | Δ vs cpu-fallback |
|---|---:|---:|---:|
| cpu-fallback (Pi 5) | 7.0 / sec | 572 ms | — |
| NPU HEF (iter 163) | **67.3 / sec** | **57 ms** | **9.6×** |
| cache hit (in-process) | **15.86 M / sec** | <1 µs | **226,000×** |
Saturation test (iter 170): 60s burst at C=100 — Pi never OOMs,
worker RSS stable at 91 MB, tonic-level backpressure correctly
drops excess requests with `ResourceExhausted`.
### Documentation
- **ADR-167** updated: NPU is now production-default
- **ADR-173** updated: ruvllm-hailo upstream embedding seam is
NPU-accelerated transparently
- **ADR-175** Option A status flipped to "production default"
- **ADR-176** new EPIC tracking the full integration arc
- **ADR-177** new — Pi 4 / no-HAT deploy
- Cluster README now has an Operator QUICKSTART covering all
three deploy paths
### Internal
- iter-153 `acceleras` Keras-registration monkey-patch — the
breakthrough that unblocked the entire HEF compile pipeline.
`compile-encoder-hef.py` walks every module under
`hailo_model_optimization.acceleras` at import time and applies
`keras.saving.register_keras_serializable()` to every
`keras.layers.Layer` subclass it finds.
## [2.0.5] - 2026-02-26
### Fixed
- **ruvector-gnn**: Replace `assert!()` with `Result` in `MultiHeadAttention::new()` and `RuvectorLayer::new()` — prevents fatal `abort()` in NAPI-RS/WASM bindings ([#216](https://github.com/ruvnet/ruvector/issues/216))
- **ruvector-gnn**: Fix pre-existing `mmap.rs` test compilation error (`grad_offset` returns `Option`)
- **install.sh**: Remove stale hardcoded version pins (`@0.1.2`, `@0.1.23`), always fetch latest
- **install.sh**: Fix operator precedence bug in CLI install guard (`--npm-only` now correctly skips CLI)
- Docs: Fix stale capability counts in root README
- Docs: Update guides to match current API surface and versions
### Added
- OpenFang Agent OS RVF example — 24 RVF capabilities demonstrated
- OpenFang project research document
- Missing capabilities added to advanced features guide
### Security
- **SEC-001**: Harden mmap pointer arithmetic with checked bounds
- **SEC-002**: Cryptographic hash binding for proof attestations (prevents spoofing)
### Changed
- Workspace version bumped from 2.0.4 to 2.0.5
- `@ruvector/gnn` bumped from 0.1.24 to 0.1.25 (all 7 platform packages)
- All WASM/NAPI wrappers (`ruvector-gnn-wasm`, `ruvector-gnn-node`, `ruvector-attention-unified-wasm`) now propagate layer construction errors as catchable JS exceptions instead of process crashes
### Published
- `ruvector-core@2.0.5` → crates.io
- `ruvector-gnn@2.0.5` → crates.io
- `@ruvector/gnn@0.1.25` → npm (linux-x64-gnu, linux-x64-musl, linux-arm64-gnu, linux-arm64-musl, darwin-x64, darwin-arm64, win32-x64-msvc)
## [2.0.4] - 2026-02-25
### Added
- **ADR-043: External Intelligence Providers** for SONA learning — pluggable external AI intelligence integration
- **Intelligence module** in `@ruvector/ruvllm@2.5.0`
- **Security Hardened RVF v3.0** — 30 verified capabilities, AIDefence + TEE hardened container (ADR-042)
- **Proof-gated graph transformer** with 8 verified modules ([#212](https://github.com/ruvnet/ruvector/pull/212))
- **Formal verification** with lean-agentic dependent types ([#206](https://github.com/ruvnet/ruvector/pull/206))
- **WASM cognitive stack** — canonical min-cut, spectral coherence, container orchestration, cold-tier GNN training ([#201](https://github.com/ruvnet/ruvector/pull/201))
- **rvDNA health biomarker analysis engine**:
- 20-SNP panel with streaming simulation
- LPA cardiovascular SNPs from SOTA meta-analysis
- CUSUM changepoint detection, gene-biomarker correlations
- SNP weights calibrated from clinical meta-analyses
- npm `@ruvector/rvdna` package with risk scoring and benchmarks
- SPARQL parser backtrack fix and executor memory leak fix in `ruvector-postgres@2.0.4`
### Security
- **Harden intelligence providers** — type-safe enums, input validation, file size limits
- **Fix path traversal** in MCP server `vector_db_backup` (CWE-22) ([#211](https://github.com/ruvnet/ruvector/pull/211))
- **Harden MCP servers** against command injection, CORS bypass, and prototype pollution ([#213](https://github.com/ruvnet/ruvector/pull/213))
### Fixed
- Migrate attention/dag/tiny-dancer to workspace versioning
- Fix all dependency version specs for crates.io publishing
- Include prebuilt binaries in `@ruvector/gnn` platform packages ([#195](https://github.com/ruvnet/ruvector/issues/195))
- CI: Node.js upgraded to 20 in GNN build workflow
- CI: Auto-publish on push to main for GNN packages
- RVF `NodeBackend` string ID ↔ numeric label mapping
## [0.3.0] - 2026-02-21
Major release introducing the RuVector Format (RVF) cognitive container, AGI runtime substrate, and a significant expansion of the platform from vector database to cognitive computing framework.
### Added
#### RuVector Format (RVF) — Universal Cognitive Container
- Complete RVF SDK with cognitive container specification ([#166](https://github.com/ruvnet/ruvector/pull/166))
- New crates: `rvf-types`, `rvf-crypto`, `rvf-runtime`, `rvf-node`, `rvf-wasm`, `rvf-solver`, `rvf-solver-wasm`, `rvf-cli`
- WASM segment (`WASM_SEG 0x10`) for self-bootstrapping RVF files
- Ed25519 asymmetric signing (RFC 8032) behind feature gate
- Witness auto-append, CLI verification, prebuilt fallbacks
- Integration into `npx ruvector` and `rvlite` (ADR-032)
- Platform-specific scripts for Linux, Windows, Node, browser, Docker
- Real Linux 6.8.12 kernel embedded in RVF for live-boot proof
#### AGI Cognitive Container (ADR-036)
- `authority_config` and `domain_profile` TLV support
- Authority guard, coherence monitor, benchmarks
- Multi-dimensional IQ with cost/robustness/AGI contract
- 5-level superintelligence pathway engine
- KnowledgeCompiler Strategy Zero, StrategyRouter bandit, ablation protocol
- Three-class memory, loop gating, RVF artifacts, rollback witnesses
- Thompson Sampling two-signal model, speculative dual-path, constraint propagation
#### QR Cognitive Seed (ADR-034)
- Pure-Rust QR code encoder for RVF seed bytes
- In-browser RVF seed decoder PWA
- Swift App Clip skeleton for iOS mobile FFI
#### Progressive Indexing Hardening (ADR-033)
- `QualityEnvelope`, triple budget caps, selective scan, fuzz benchmark
- `ResultQuality` extended to API boundary
- Malicious manifest test and brute-force cap
#### Sublinear-Time Sparse Solver
- Complete `ruvector-solver` crate with zero-overhead SpMV
- Fused Neumann iteration kernel
- WASM solver: self-learning AGI engine compiled to WASM
- Min-cut gating experiment modules
#### Additional Systems
- **RvBot**: Self-contained RVF bot with real Linux 6.6 kernel and initramfs boot
- **rvDNA Genomics**: Complete SOTA genomic analysis pipeline, native 23andMe genotyping v0.2.0
- **Domain Expansion**: Cross-domain AGI transfer learning engine with WASM bindings and meta-learning
- **OSPipe**: RuVector-enhanced personal AI memory for Screenpipe ([#163](https://github.com/ruvnet/ruvector/pull/163))
- **Quantum Simulation**: `ruqu-core`, `ruqu-algorithms`, `ruqu-wasm`, Bell test CHSH inequality
- **Causal Atlas** (ADR-040): Dashboard, solver, and desktop app
- **ruvector-postgres v0.3.0**: 43 new SQL functions (ADR-044)
### Fixed
- HNSW index bugs, agent/SPARQL crashes ([#152](https://github.com/ruvnet/ruvector/issues/152), [#164](https://github.com/ruvnet/ruvector/issues/164), [#167](https://github.com/ruvnet/ruvector/issues/167), [#171](https://github.com/ruvnet/ruvector/issues/171))
- LRU security fix ([#148](https://github.com/ruvnet/ruvector/issues/148))
- FPGA-transformer `BackendSpec.as_ref` and HNSW array indexing
- Platform-specific errno on macOS/BSD ([#174](https://github.com/ruvnet/ruvector/issues/174))
- WASM path resolution in CJS→ESM interop
- Docker Rust version bumped to 1.85 for edition2024
### Changed
- `rvf-types`, `rvf-crypto`, `rvf-runtime` bumped to 0.2.0
- npm: `ruvector@0.1.99`, `rvlite@0.2.4`, `rvf@0.1.3`
## [0.2.6] - 2025-12-09
### Added
- **`ruvector-postgres` PostgreSQL extension** with SIMD optimizations and 53 SQL function definitions
- **PostgreSQL 18 support** with backward compatibility
- **`@ruvector/postgres-cli`** with native installation support
- **W3C SPARQL 1.1 query language** support in PostgreSQL extension
- **GNN v2** comprehensive implementation with cognitive substrate
- **iOS-optimized WASM recommendation engine**
- **9 cognitive substrate crates** published as EXO-AI 2025
- **Neuromorphic HNSW v2.3** with SNN (Spiking Neural Network) integration
- **Ultra-low-latency meta-simulation engine** example
- **8 specialized Docker images** with publishing infrastructure
- **RuVector Studio** — complete web UI application
- `ruvector-attention` functions exported from PostgreSQL extension
### Fixed
- Docker build and extension SQL for PG17
- SPARQL build compilation — achieved 100% clean build
- Docker Hub README and image references
### Changed
- npm packages reorganized from `/src` to `/npm/packages`
### Breaking Changes
- npm import paths changed due to `/src` → `/npm/packages` reorganization
## [0.1.32] - 2026-01-17
### Added
- **SONA Neural Architecture** npm package (`sona@0.1.5`)
- **RuvLLM** npm package with intelligence module
- **Graph Node** bindings (`@ruvector/graph-node@0.1.26`)
- npm package expansion and version consolidation
## [0.1.19] - 2025-12-01
### Fixed
- **GNN Node.js bindings**: Use `Float32Array` for NAPI bindings to fix type conversion errors
## [0.1.16] - 2025-11-27
### Added
- **Persistent GNN layer caching** — 250-500x performance improvement
- **Self-learning GNN strategy** for accuracy improvement
- GNN NAPI-RS bindings for all platforms
## [0.1.0] - 2025-11-25
Initial release of RuVector — a high-performance vector database written in Rust.
### Added
#### Core Vector Database
- HNSW (Hierarchical Navigable Small World) graph indexing
- SIMD-optimized distance metrics (Euclidean, Cosine, Dot Product, Manhattan)
- Memory-mapped vector access via memmap2
- Parallel index construction using rayon
- Zero-copy serialization with rkyv
- Scalar quantization (int8) for 4x memory compression
#### AgenticDB Compatibility Layer
- Full 5-table schema: `vectors_table`, `reflexion_episodes`, `skills_library`, `causal_edges`, `learning_sessions`
- Reflexion Memory API with semantic search over self-critique episodes
- Skill Library with auto-consolidation and usage tracking
- Causal Memory Graph with confidence scoring and hypergraph support
- 9 RL algorithms (Q-Learning, SARSA, DQN, PPO, Actor-Critic, Policy Gradient, Decision Transformer, MCTS, Model-Based)
#### Advanced Search
- Product Quantization (PQ) with 8-16x memory compression at 90-95% recall
- Filtered search (pre/post-filtering with complex expressions)
- Hybrid search (vector similarity + BM25 keyword scoring)
- MMR (Maximal Marginal Relevance) diversity-aware ranking
- Conformal prediction with distribution-free confidence intervals
#### Multi-Platform Deployment
- **Node.js** (NAPI-RS): Async API, TypeScript types, zero-copy Float32Array
- **WASM**: Browser-compatible, Web Workers, IndexedDB persistence
- **CLI**: JSON/CSV/NPY support, shell completions, benchmarking
- **Cross-platform builds**: Linux (x64/arm64), macOS (x64/arm64), Windows (x64), WASM
#### Performance
- 10-100x faster than Python/TypeScript implementations
- Sub-millisecond latency (p50 < 0.8ms for 1M vectors)
- 95%+ recall with HNSW (ef_search=100)
- 4-32x memory compression with quantization
- 200-300x distance calculation speedup with SIMD
- Near-linear scaling to CPU core count
### Dependencies
- **Core**: redb, memmap2, hnsw_rs, simsimd, rayon, crossbeam
- **Serialization**: rkyv, bincode, serde, serde_json
- **Node.js**: napi, napi-derive
- **WASM**: wasm-bindgen, wasm-bindgen-futures, js-sys, web-sys
- **Math**: ndarray, rand, rand_distr
- **CLI**: clap, indicatif, console
---
For questions or issues, visit: https://github.com/ruvnet/ruvector/issues