RetrievalKit v0.1.0 preview

Changelog

All notable user-facing changes and persistence migrations are recorded here.

0.1.0 - Unreleased preview

Added

• Optional v0.1.0 embedding packages for Python
  (retrievalkit-embedding), Node
  (@gungorbasa/retrievalkit-embedding), browser Workers
  (@gungorbasa/retrievalkit-browser-embedding), Kotlin/JVM
  (io.github.gungorbasa:retrievalkit-embedding), and Android arm64-v8a
  (io.github.gungorbasa:retrievalkit-embedding-android). Rust embedding
  crates remain source-only.
• TypeScript/Node base and graph aggregates for Node.js LTS on macOS arm64,
  selected for publication as @gungorbasa/retrievalkit and
  @gungorbasa/retrievalkit-graph after npm rejected the equivalent unscoped
  base name as too similar to an existing package.
  Promise-based N-API operations use typed values, Float32Array, exact
  bigint transport, Rust-owned search and graph semantics, deterministic
  async disposal, package-content checks, and local-install smoke tests.
• Browser retrieval and graph support as the independent
  @gungorbasa/retrievalkit-browser npm package. Its dedicated Worker owns the
  in-memory database, and its release tarball includes separately qualified
  portable and SIMD128 WebAssembly tiers; browser embedding remains a separate
  optional package.
• Kotlin/JVM and Android base and graph aggregates under the selected
  io.github.gungorbasa Maven group, with typed JNI transport, FloatArray
  embeddings, AutoCloseable lifecycle, opaque synchronized handles, and
  Android arm64-v8a AAR packaging. Android API 24+ arm64-v8a ships as an
  explicit preview: host-verifiable build, package, inventory, ABI/JNI, and
  fresh consumer compilation/resolution checks remain required, while live
  device inference, compatibility, and performance are unqualified, deferred,
  and non-blocking for v0.1.0 publication.
• Python graph queries, results, and stable candidate projection now cross
  PyO3 as typed values without JSON. Projection filtering, stale-selection
  checks, ordering, and counts remain owned by the canonical Rust corpus.
• Progressive Python, TypeScript, and Kotlin ingestion accepts ordinary
  documents plus direct embeddings and infers dimension in Rust, while
  preserving advanced compatibility APIs where they already existed.
• Crash-safe transactional index saves. New data is written to an immutable
  generation and synced before manifest.json publishes it.
• Cross-process save locking prevents concurrent writers from publishing and
  cleaning generations out of order. Locks are released by the operating system
  if a process exits or crashes.
• Explicit Rust, Swift, and Python index compaction removes tombstoned payloads,
  preserves active chunk IDs, reports estimated reclaimed memory, and remains a
  cheap no-op when there is nothing to reclaim. Compaction is a synchronous
  maintenance operation and temporarily retains old and replacement structures
  to guarantee an all-or-nothing swap.
• Checksummed persistence format V4 verifies vectors, chunks, canonical records
  and their stable external/internal chunk mappings, BM25, and tombstones with
  SHA-256 before loading. Rust, Swift, and Python expose read-only validation
  APIs and typed corruption failures.
• Parallel Swift exact, keyword, and hybrid searches on one VectorIndex, with
  writer-preferring exclusive access for upsert, delete, save, and compaction.
  The C/FFI threading and handle-lifetime contract is now explicit.
• Python retrieval, persistence, and maintenance release the GIL during
  Rust-only work. Shared-index searches may run across Python threads, while
  PyO3 exclusive borrowing rejects conflicting mutation safely.
• An isolated memory benchmark now measures sampled peak RSS across build,
  cold/warm search, save, unload, load, delete, and compaction. JSON budgets can
  fail the CLI, and the iOS app provides one-scenario-per-launch target presets.
• Automated iOS memory presets print JSON to the attached device console and
  exit with a budget-aware status. iPhone 17 Pro Max measurements now define
  provisional 24K compact-target and 50K extended-capacity budgets.
• A versioned MiniLM retrieval-quality fixture now gates relevance, F32/I8
  overlap, candidate limits, filters, deletions, replacements, persistence
  reload, and latency. Its V1 evidence keeps 50/50 as the hybrid default and
  confirms 98.33% top-5 and 100% top-10 I8/F32 vector-only overlap.
• New Rust, Swift, and Python indexes now default to I8 scalar-quantized vector
  storage. Hybrid queries default to 50/50 candidates and weighted normalized
  score fusion with query-time alpha = 0.6, matching the public search API.
  F32, different alpha values, and explicit candidate limits remain available.
• Retrieval-quality V2 expands the benchmark to 306 documents and 42 graded,
  ambiguous queries. It adds a human relevance-recall gate while preserving V1
  as a historical baseline.
• A manual-only verification workflow can check Rust, the V2 retrieval-quality
  gates, Python typing/lint/tests and installed wheels, plus Apple XCFramework
  and Swift builds without running automatically or publishing artifacts.
• Evidence-led README with claim-register enforcement and tested
  capability-oriented quickstarts.
• Combined Swift/Python release-candidate tooling with reproducible artifact
  inventory, SBOM, provenance, and fail-closed publication authorization.
• Token-aware Swift and Python ingestion pipelines with custom chunker support.
• Clean-source onboarding qualification now measures Python, Swift, Node.js,
  and Kotlin time to first result, records an explicit evidence schema and
  Swift toolchain metadata, and runs monthly or on demand.
• Source-preview documentation now includes searchable Swift guidance,
  self-contained language examples, responsive mobile actions, and a recovery
  page for unknown routes.
• Node.js wrapper tooling now accepts the maintained Node.js 22.13+ and 24 LTS
  ranges, keeps package engine declarations synchronized, and has zero npm
  audit findings after its development-tool refresh. Kotlin preflight now
  distinguishes the required JDK 17 build toolchain from the Java 11 bytecode
  target and reports the exact selected Java binary with recovery commands.
• Fail-closed Node and Kotlin release assemblers now produce inspected,
  deterministic macOS arm64 npm tarballs and JVM/Android Maven artifacts using
  the exact selected registry identities. The two-root release-candidate
  workflow includes those artifacts in the same closed Swift/Python/Node/Kotlin
  bundle and rejects any alternate npm name or Maven group.
• Publication authorization is now created at runtime after a required reviewer
  approves the protected GitHub release environment. The resulting provenance
  record binds the signed tag, source revision, workflow runs, Phase 7 results,
  candidate inventory, and checksums without requiring an impossible
  self-referential authorization file in the release commit.
• Guarded npm and Maven Central publication jobs consume the authorized
  candidate bytes without rebuilding them. npm uses trusted publishing with
  provenance; Maven signs the exact primary files with the dedicated
  RetrievalKit release key before uploading a Central Portal bundle. All five
  scoped npm names now have bootstrap-only placeholders and trust the
  protected GitHub publication workflow. All three PyPI names now have non-SDK
  0.0.0a0 placeholders and trust the same protected workflow through the
  pypi environment. The io.github.gungorbasa Central namespace is verified,
  its signing identity is public, and all five Maven publishing secrets are
  installed in the protected environment. v0.1.0 remains unpublished; the
  signed tag and provisioned release gates remain external prerequisites.

Compatibility

• Search result buffers now use a packed UTF-8 arena and offset/length ranges
  instead of separately allocated C strings. Effective result metadata uses a
  flat packed entry table referencing the same arena, and hybrid alpha is
  stored once per buffer. Native boundary types and constants now use the
  product-aligned RetrievalKit/RETRIEVALKIT_ prefixes instead of the stale
  pre-rename Vk/VK_ prefixes. The graph aggregate ABI version is 12. Native
  libraries, headers, and wrappers from different ABI versions must not be
  mixed.
• Public hybrid alpha endpoints now disable candidate generation from the
  zero-weight source: alpha = 1 is truly vector-only and alpha = 0 is truly
  BM25-only. Generic C fusion/RRF entrypoints were removed; RRF remains an
  internal Rust benchmark option.
• Swift and Python exact, BM25, and hybrid hits now share the canonical result
  metadata contract. Hybrid traces expose alpha; the constant
  filterMatched/filter_matched field and public fusion dictionaries were
  removed. Swift graph metadata now uses the shared MetadataValue;
  GraphMetadataValue remains as a deprecated compatibility alias.
• Swift now publishes RetrievalKit and RetrievalKitGraph from one package
  backed by one graph-capable native aggregate. Applications may select either
  or both products without linking competing native libraries. TextChunker
  is part of RetrievalKit; the separate RetrievalKitIngest product was
  removed.
• Invalid hybrid alpha values are query-argument errors in Rust, Swift, and
  Python rather than being mislabeled as an invalid persisted index format.
• RetrievalKitPipeline now accepts the shared typed DocumentID used by the
  progressive Swift API while preserving its existing string result surface.
• V1, V2, and V3 indexes remain readable; their next save publishes a
  checksummed V4 snapshot under .snapshots.
• Index directories should be treated as RetrievalKit-owned. Applications must
  not modify .snapshots or manifest.json directly.

Upgrade

Rebuild and upgrade the Rust native artifact, C headers, a...