kalan_db 0.1.0

Offline hybrid search for Flutter (Android). HNSW + BM25, RRF-fused, with on-device ONNX query embedding — no server required.

0.1.0

Initial public release of Kalan DB.

• Embedded, offline hybrid search: HNSW vector + BM25 keyword, fused with Reciprocal Rank Fusion (configurable alpha).
• On-device text embedding via ONNX Runtime (flutter_onnxruntime): ship a small int8 model and embed queries at runtime — no server, no pre-computed query vectors.
  • TextEmbedder with a pure-Dart BERT WordPiece tokenizer (validated against the HuggingFace reference), CLS/mean pooling, L2 normalization.
  • EmbeddedKalanDB convenience wrapper: search(String text) embeds then searches in one call.
  • Tuned defaults for bge-small-en-v1.5 (CLS pooling + query instruction); configurable for e5 / all-MiniLM (mean pooling).
• Memory-mapped, read-only .vdb format with header + xxHash64 integrity check.
• Metadata pre-filtering (subject, grade, chapter, language, book id).
• Python offline builder (PDF → chunks → embeddings → .vdb) and a native vdb_build / vdb_query host tool.
• On-device KPI harness (example): Precision@5, Recall@10, NDCG@10, MRR plus embed/search latency, evaluated against Gemini-graded relevance judgments.
• Android (arm64-v8a) only in this release.