Aligns the C# port's query / extraction / storage behavior with Python LightRAG and ships a single merged assembly.
Highlights
- Query: graph 1-hop expansion (local→relations, global→endpoint entities), weighted-polling / vector chunk selection, rerank + min_rerank_score + chunk_top_k, Python-compact context JSON, dynamic chunk token budget
- Extraction: record + tuple-delimiter recovery, 256-char / 512-byte entity-name limit, 3 JSON-mode few-shot examples
- KG merge: map-reduce summarization (no description loss), empty-relation raises
- Storage: doc-status immediate flush, VDB meta_fields whitelist, typed GraphML, degree-priority subgraph, DocStatus enum + zero-seeded counts
- Caching: flattened, model-partitioned cache key
- Single
LightRAG.NET.dll(Core + Storage.FileBased, netstandard2.1, Unity-ready). Tests 41 → 64.
Resolves all 18 fidelity gaps from the parity review (5 critical, 13 major).
Assets
LightRAG.NET.dll— the merged managed plugin (netstandard2.1).LightRAG.NET-v0.2.0-netstandard2.1.zip— self-contained bundle: the DLL plus its runtime dependencies (Newtonsoft.Json, Microsoft.ML.Tokenizers + data, BCL/System shims). OllamaSharp is not required (the Ollama provider is excluded from this build).