…flags

…ne, Euclidean, VectorOps)

…p) with zero-copy I/O

…rdAnalyzer

…l threads

…gestion pipeline

…mark scaffold

… SECURITY, README, CI, templates)

… HNSW recall tests

…to-detect

…n-testtools dependency

…25 scoring

…hunker, TextUtils and add chunked ingestion for large documents

…kenizer for large document support

… auto-embed engine integration

- QuantizationType enum (NONE, SCALAR_INT8)
- ScalarQuantizer with min/max calibration and INT8 encoding
- QuantizedCosineSimilarity and QuantizedDotProduct SIMD kernels
- SimilarityFunction updated with quantized variants
- ScalarQuantizerTest for encode/decode and batch operations

- PersistenceMode enum (IN_MEMORY, DISK, MMAP)
- IndexFileFormat for binary HNSW serialization
- QuantizedVectorStore with INT8 compression
- InMemoryVectorStore concurrent access improvements

- DiskHnswWriter for binary HNSW graph serialization
- DiskHnswIndex for mmap-based read-only index loading
- QuantizedHnswIndex with INT8 scalar quantization (4x memory reduction)
- BM25Index and HnswIndex performance improvements
- DiskHnswIndexTest and QuantizedHnswIndexTest

- ProductQuantizer: K-Means++ codebook training, PQ encode/decode, ADC
  distance computation, batch encoding
- IvfPqIndex: full IVF-PQ implementing VectorIndex SPI with cluster
  assignment, residual-based PQ encoding, and multi-probe search
- PostingList: per-cluster growable storage for PQ codes
- 14 tests: PQ training/encode/decode/ADC + IVF-PQ search/recall/sorting

- Reranker SPI interface for pluggable re-ranking strategies
- LlmReranker: listwise relevance scoring using Ollama generate API
  with prompt-based 0-10 scoring and graceful fallback
- HybridSearchOrchestrator: integrated optional re-ranking post-processing
- LlmRerankerTest: fallback behavior, empty input, topK limiting

- spector-gpu Maven module with Panama FFM CUDA bindings
- GpuCapability: runtime CUDA detection (device count, name, memory)
- GpuBatchSimilarity: SIMD-optimized batch dot product and cosine
  similarity using FMA Vector API operations
- CudaKernelLauncher: PTX module loader, function resolver, kernel
  launcher with grid/block configuration
- batch_similarity.cu: CUDA kernels for batch_cosine, batch_dot, batch_l2
  with block-level shared memory reduction
- 14 tests: GPU detection, batch similarity correctness, CUDA launcher

…hitecture

- spector-cluster Maven module with gRPC/protobuf integration
- spector_search.proto: 6 RPC definitions (vector, keyword, hybrid
  search, ingest, health check, stats)
- ClusterCoordinator: fan-out/merge query execution via virtual threads
  with consistent hash shard routing
- ShardNode: gRPC server wrapping SpectorEngine
- SpectorSearchServiceImpl: full gRPC service delegating to local engine
- RemoteShardClient: type-safe gRPC client for all 5 RPC methods
- ClusterConfig: multi-node endpoint configuration with replication
- ClusterConfigTest: routing, hash consistency, topology tests

…rEngine

- IndexType enum (HNSW, IVF_PQ) for configurable index strategy
- SpectorConfig: added indexType, ivfNlist, ivfNprobe, pqSubspaces with
  builder methods (withIvfPq) and auto-defaults
- SpectorEngine: IVF-PQ auto-training pipeline that buffers ingested
  vectors and trains PQ codebooks after nlist*40 samples
- Backward-compatible 7-arg constructor preserved
- 4 new tests: auto-training, keyword search during buffering, config
  builder, auto-defaults

- HeavyPerformanceBenchmark: keyword/vector/hybrid at 50K-100K scale
- IvfPqBenchmark: IVF-PQ search, PQ encode/decode, ADC distance, batch
  cosine similarity at 10K-50K scale
- ConcurrencyBenchmark: multi-threaded search throughput
- IngestionBenchmark: document ingestion throughput
- PerformanceTestRunner: standalone runner with formatted results

- pom.xml: added spector-gpu, spector-cluster modules to reactor
  and dependencyManagement
- README.md: expanded architecture (13 modules), 5 new features,
  updated comparison table (quantization, IVF-PQ, GPU, LLM, distributed),
  updated test suite (316+ tests), added roadmap checklist

Extract ~300 lines of duplicated graph traversal code (greedyClosest,
searchLayer, selectNeighbors, addConnection, getNeighbors, setNeighbors)
into AbstractHnswIndex base class with three template method hooks:

- computeDistance(float[], int) — distance from query to stored node
- getNodeVector(int) — float32 vector retrieval for pruning
- storeVector(int, float[]) — vector storage on insertion

HnswIndex: 413 -> 76 lines (-81%)
QuantizedHnswIndex: 476 -> 226 lines (-53%)

All 316+ tests passing, zero regressions.

- VectorIndex: add default isReadOnly() method (returns false)
- DiskHnswIndex: override isReadOnly() to return true
- KeywordIndex: add default remove(String id) method
- BM25Index: expose existing removeDoc() logic via KeywordIndex.remove()

Completes the deletion API path across the engine.

Introduce GoF design patterns for component creation:

- VectorIndexFactory: creates HNSW, QuantizedHNSW, or IVF-PQ based on
  config (replaces if/else chain in engine constructor)
- VectorStoreFactory: creates InMemory or MappedVectorStore based on
  PersistenceMode (replaces hardcoded InMemoryVectorStore)
- EngineComponentFactory: Abstract Factory assembling all components
  (store, index, GPU, reranker) into an EngineComponents record
- EngineComponents: immutable record grouping all subsystems

Adding a new index or store type now requires zero changes to
SpectorEngine (Open/Closed Principle).

Refactor SpectorEngine to delegate component construction to
EngineComponentFactory (Abstract Factory) instead of inline if/else:

- Constructor: 150 lines -> 30 lines
- Field type: BM25Index -> KeywordIndex (DIP compliance)
- Removed 8 concrete class imports for construction (now in factories)
- Added SpectorEngine.Builder for fluent engine construction:

    SpectorEngine engine = SpectorEngine.builder()
        .dimensions(384).capacity(100_000)
        .similarity(SimilarityFunction.COSINE)
        .gpu(true).build();

- Added constructor accepting custom EngineComponentFactory for testing
- Integrated GPU fallback and LLM reranker lifecycle

…points

- CORS support via Javalin bundled plugin
- Optional API key authentication via X-API-Key header
- Vector dimension validation on ingest
- New endpoints: /api/v1/ingest/auto, /api/v1/ingest/bulk,
  DELETE /api/v1/documents/{id}, /api/v1/metrics
- Request counters via LongAdder for observability

Add 4-arg constructor accepting CA cert, client cert, and client key
for TLS-encrypted gRPC connections. Transparent fallback to plaintext
for development environments.

…ReentrantLock

- Parent POM: remove duplicate jackson-databind declaration
- MappedVectorStore: replace synchronized with ReentrantLock on put()
  and close() for virtual thread compatibility (consistent with
  InMemoryVectorStore)

- README: reflect current 13-module architecture, design patterns,
  GPU acceleration, LLM re-ranking, and IVF-PQ indexing
- CHANGELOG: comprehensive feature inventory across all modules
- goal.md: update roadmap with completed items and current status

…ernels

…dd INT4/INT2 index support

…configuration

…nd Spring AI integration

Bumps [actions/upload-artifact](https://github.com/actions/upload-artifact) from 4 to 7.
- [Release notes](https://github.com/actions/upload-artifact/releases)
- [Commits](actions/upload-artifact@v4...v7)

---
updated-dependencies:
- dependency-name: actions/upload-artifact
  dependency-version: '7'
  dependency-type: direct:production
  update-type: version-update:semver-major
...

Signed-off-by: dependabot[bot] <support@github.com>