Executive Summary
Benchmark testing of AgentDB v2.0.0 with integrated RuVector shows dramatic performance improvements compared to baseline measurements from November 29. The optimizations deliver sub-microsecond search latencies and throughput exceeding 190,000 queries per second.
Key Findings
| Metric |
Yesterday (Nov 29) |
Today (Nov 30) |
Improvement |
| Search p50 (2K vectors) |
256.3 µs |
1.5 µs |
170x faster |
| Search p99 (2K vectors) |
290.3 µs |
8.0 µs |
36x faster |
| QPS (2K vectors) |
3,638 |
192,840 |
53x higher |
| Batch insert throughput |
20,945 ops/s |
2,703,923 ops/s |
129x faster |
| Search p50 (12K vectors) |
1,618.7 µs |
2.2 µs |
735x faster |
| QPS (12K vectors) |
559 |
84,138 |
150x higher |
Test Environment
- Platform: Linux ARM64 (aarch64)
- Node.js: v22.x
- Package: agentdb@2.0.0
- RuVector: @ruvector/core (with ruvector-core-linux-arm64-gnu)
- Vector Dimension: 384
- Distance Metric: Cosine
Detailed Results
@ruvector/core Direct Benchmark
Insert Performance (1,000 vectors)
| Metric |
Value |
| Avg insert |
5.8 µs |
| Insert p50 |
3.3 µs |
| Insert p99 |
69.3 µs |
| Throughput |
171,055 ops/sec |
Batch Insert (1,000 vectors)
| Metric |
Value |
| Total time |
0.37 ms |
| Throughput |
2,703,923 ops/sec |
Search Performance (k=10, 2,000 vectors)
| Metric |
Value |
| Avg latency |
2.5 µs |
| p50 latency |
1.5 µs |
| p99 latency |
8.0 µs |
| Min latency |
0.8 µs |
| Max latency |
294.0 µs |
| QPS |
192,840 queries/sec |
Scale Test (12,000 vectors)
| Metric |
Value |
| 10K batch insert |
5.52 ms |
| Insert throughput |
1,811,854 ops/sec |
| Search p50 |
2.2 µs |
| Search p99 |
11.7 µs |
| QPS at 12K |
84,138 queries/sec |
@ruvector/graph-node Benchmark
| Operation |
Throughput |
Avg Latency |
| Single node create |
10,033 ops/sec |
0.0997 ms |
| Batch create (100) |
346,875 nodes/sec |
0.2883 ms |
| MATCH simple query |
4,760 qps |
0.21 ms |
| MATCH with WHERE |
4,997 qps |
0.20 ms |
AgentDB SDK Integration
| Operation |
Throughput |
Avg Latency |
| Store Episode |
177 ops/sec |
5.64 ms |
| Retrieve Episodes |
1,910 ops/sec |
0.52 ms |
Issues Encountered
1. @ruvector/core ARM64 Binding
- Issue: Initial load failed with "Failed to load native binding for linux-arm64"
- Solution: Installed
ruvector-core-linux-arm64-gnu package
2. Build TypeScript Errors
- Issue:
npm run build fails with 27+ TS errors in simulation/ and cli/ modules
- Impact: Pre-built dist files work, but fresh builds fail
- Recommendation: Fix TypeScript errors before release
3. GNN Module Timeouts
- Issue: @ruvector/gnn tests timeout (60s)
- Impact: GNN forward pass benchmarks incomplete
- Recommendation: Review GNN initialization performance
4. SkillLibrary SQLite Fallback
- Issue:
this.db.prepare is not a function when using GraphDatabase
- Impact: SkillLibrary benchmark fails
- Recommendation: Ensure consistent API between GraphDB and SQLite modes
Recommendations
For Development
- Fix TypeScript Build: 27 compilation errors prevent clean builds
- GNN Module: Investigate timeout issues on ARM64
- SDK Integration: Ensure SkillLibrary works with GraphDatabase mode
- Add ARM64 to CI: Include ARM64 in test matrix
For Integration
- Adopt AgentDB v2: Performance gains are substantial
- Use @ruvector/core directly: Bypass SDK overhead for hot paths
- Batch operations: 129x improvement justifies batching strategy
- ARM64 Support: Native bindings work well, recommend as supported platform
Reference
Conclusion
AgentDB v2.0.0 with RuVector integration delivers exceptional performance improvements:
- 170x faster search latency at 2K scale
- 735x faster search latency at 12K scale
- 129x faster batch inserts
- 53-150x higher throughput (QPS)
Report generated by Agentic QE Fleet
Benchmark date: 2025-11-30
