taerae_core #

taerae_core is an embedded graph database core for Dart. It provides an in-memory graph engine, file-backed persistence, and GraphRAG integration points.

Highlights #

• Immutable models: TaeraeNode, TaeraeEdge
• In-memory graph APIs: TaeraeGraph
• File persistence with log + snapshot: TaeraePersistentGraph
• GraphRAG components: TaeraeGraphRag, TaeraeTextEmbedder, TaeraeVectorIndex

Core Differentiators #

• Embedded-first: no external graph server dependency for core graph features.
• Local-first durability: append-only log + snapshot with configurable durability policy.
• Dart-native API: no query language bridge required for common graph operations.
• GraphRAG extension ready: structured hooks for embedding, indexing, filtering, and reranking.

Quick start #

import 'package:taerae_core/taerae_core.dart';

void main() {
  final TaeraeGraph graph = TaeraeGraph()
    ..upsertNode('alice', labels: const <String>['Person'])
    ..upsertNode('bob', labels: const <String>['Person'])
    ..upsertEdge('e1', 'alice', 'bob', type: 'KNOWS');

  final List<String>? path = graph.shortestPathBfs('alice', 'bob');
  print(path); // [alice, bob]
}

Detailed guide #

• Developer Guide Detailed usage guide covering CRUD, traversal, JSON import/export, persistence durability, recovery strategy, GraphRAG workflows, defensive coding, performance, and testing patterns.

Benchmark #

Measure core performance for larger datasets with the bundled benchmark script.

dart run benchmark/graph_search_benchmark.dart --preset=generic

Workload presets:

• generic: balanced indexed graph workload.
• social: user follow/recommendation pattern (FOLLOWS, HAS_INTEREST).
• delivery: route + package tracking pattern (ROUTE, DELIVERS_TO).
• notes_rag: note-link + filter-heavy pattern (RELATED, TAGGED_AS).

Example runs:

dart run benchmark/graph_search_benchmark.dart \
  --preset=social \
  --sizes=20000,100000 \
  --path-queries=300

dart run benchmark/graph_search_benchmark.dart \
  --preset=delivery \
  --sizes=20000 \
  --path-queries=500

The benchmark reports build throughput (upsertNode, upsertEdge) and search performance (label/property lookup, adjacency traversal, and shortest-path style queries) for each dataset size.

For paper-grade reporting (repeated runs + statistics), use the runner:

dart run benchmark/paper_benchmark.dart \
  --presets=social,delivery,notes_rag \
  --sizes=20000,100000 \
  --warmup-runs=1 \
  --repeat=5

It writes:

• benchmark/results/<timestamp>/results.json
• benchmark/results/<timestamp>/summary.csv
• benchmark/results/<timestamp>/REPORT.md