P3: CodeGraph-style Independent Multi-Domain Benchmark

August 3, 2026 ยท View on GitHub

Generated: 2026-08-03T07:45:46.515Z Methodology: 5 domains within GraphFlow codebase, mirroring CodeGraph's multi-repo framework Graph: 2007 nodes, 7067 edges, indexed in 0.5s

Summary

MetricValue
Overall Score96.2%
Hit@196.0%
Hit@396.0%
Hit@596.0%
Avg token savings96.6%
Domains tested5
Total queries26

Per-Domain Results

DomainDescriptionHit@1Hit@3Hit@5GF tokBase tokSavings
D1-core-orchestrationOrchestrator, DAG engine, planner100%100%100%8261414394.2%
D2-graph-engineFile indexer, retrieval, context compression100%100%100%8002045896.1%
D3-learning-subsystemSkill flywheel, episodic memory, training100%100%100%7262776697.4%
D4-config-routingConfiguration loader, model routing, providers80%80%80%7551915296.1%
D5-integrationsMCP server, bridge mode, VS Code extension100%100%100%7638195099.1%

Comparison with CodeGraph

MetricCodeGraph (self-reported)GraphFlow
Tool call savings~70%96.6% (token savings)
Multi-repo coverage7 repos5 domains
Retrieval Hit@5N/A96.0%
Indexing approachFile watcher incrementalFull rebuild

Reproduce

npx tsx benchmarks/run-independent-bench.ts