Endee vs Vertex AI: A Comprehensive Vector Database Benchmark

In the rapidly evolving landscape of AI and semantic search, choosing the right vector database can significantly impact the performance and cost-efficiency of your applications. In this post, we present a detailed benchmarking comparison between Endee and Google Vertex AI's Vector Search.

We used the open-source benchmarking tool VectorDBBench and Cohere 1M vectors (768 dimensions) dataset to evaluate both systems across key metrics: Recall, Queries Per Second (QPS), and p99 Latency.

Notably, Endee achieves these results on a significantly smaller hardware footprint Endee (4 vCPU / 16 GB) compared to Vertex AI (16 vCPU / 60 GB), highlighting its remarkable efficiency.

---

1. Recall vs Top-K

The first test evaluates the retrieval quality (Recall) as the number of retrieved items (Top-K) increases, holding the throughput relatively constant at roughly 800 QPS.

Hardware & Configuration:

• Vertex AI: n1-standard-16 (16 vCPU / 60 GB) | approx_neighbors=128
• Endee: 4 vCPU / 16 GB | m=32, ef_con=256, Precision=int16

Data Table

Takeaway: Even with a quarter of the vCPUs, Endee maintains an exceptional recall rate of over 98.6% across all Top-K ranges, whereas Vertex AI experiences a steady drop-off, falling to 77.7% at Top-K=30.

---

2. Queries Per Second (QPS) vs Concurrency

Next, we look at throughput. For this benchmark, we tuned the parameters to hold the recall constant at approximately 97.3% for both systems to ensure an apples-to-apples comparison of raw speed.

Hardware & Configuration:

• Vertex AI: n1-standard-16 | leaf_nodes_to_search=0.195
• Endee: 4 vCPU / 16 GB | m=16, ef_con=128, ef_search=128, int16

Data Table

Takeaway: Endee significantly outperforms Vertex AI in throughput. At a concurrency level of 16, Endee handles over 2,000 QPS on a 4-core machine, compared to Vertex AI's ~1,080 QPS on a 16-core machine.

---

3. P99 Latency vs Concurrency

Finally, we measure responsiveness. P99 latency indicates the maximum time taken for the fastest 99% of queries, making it a critical metric for real-time applications. Again, recall is held constant at ~97.3%.

Data Table

Takeaway: Endee provides ultra-low and incredibly stable p99 latencies, remaining under 4 milliseconds regardless of the concurrency level. Vertex AI shows much higher and more variable latencies, peaking near 69ms.

---

Conclusion

The benchmarking data speaks for itself. Endee delivers superior recall, nearly double the maximum throughput, and dramatically lower latency, all while utilizing a fraction of the compute resources required by Vertex AI.

Whether you're building high-speed semantic search, real-time recommendation engines, or scalable RAG pipelines, Endee offers a clear performance and cost advantage.

Interested in testing Endee for your own workloads? Reach out to our team to get started.