Benchmarks.
Fast and accurate retrieval is the backbone of the new generation of AI-native applications.
TopK is built for teams working with large volumes of data where high performance, flexibility, and cost efficiency matter.
01
Enterprise retrieval
Compare topk-embed-v1 with managed embedding providers on eight public ViDoRe v3 datasets, using page images or Markdown.
ViDoRe v3 · managed providers · upper left is better
View chart data and sources
| Managed provider | USD / 1,000 images | nDCG@10 |
|---|---|---|
| topk-embed-v1 · production | $0.075 | 66.33 |
| Cohere Embed 4 | ≈$2.40 | 58.23 |
| Voyage Multimodal 3.5 | $0.60 | 57.54 |
| Gemini Embedding 2 | $0.12 | 56.00 |
| Mixedbread Wholembed V3 | $3.00 | 64.24 |
nDCG@10 on a 0–100 scale, averaged equally across eight public ViDoRe v3 datasets. Native uses each dataset’s original query language; multilingual includes all six languages, including native.
TopK: $0.075/1,000 images. Gemini: $0.00012/image. Mixedbread: 2,000 content tokens/image at $1.50/M. Voyage: one-megapixel images at $0.60/billion pixels. Cohere is an estimate for one-megapixel images, applying AWS’s token formula (pixels ÷ 784 × 4) to Cohere’s $0.47/M image-token rate; These are billing estimates, not measured benchmark bills.
Rates: Google, Cohere, Voyage, Mixedbread (September 24, 2026). Mixedbread’s $1.50/M rate includes indexing.
02
Agentic search
Compare answer accuracy and search calls per question on BrowseComp-Plus with GPT-5 as the agent.
BrowseComp-Plus · GPT-5 agent · upper left is better
View chart data and sources
| Measurement | Tool calls per question | Accuracy (%) |
|---|---|---|
| Mixedbread | 11.53 | 90.48 |
| TopK (topk-embed-v1) | 11.8 | 86.87 |
| qwen3-embed-8B (dense) | 21.74 | 71.69 |
| BM25 (sparse) | 23.23 | 57.59 |
BrowseComp-Plus results with GPT-5 as the agent. Accuracy is the percentage of correctly answered questions; search calls are averaged per question.
03
Billion documents in milliseconds
Predictable latency at scale with filtered or unfiltered queries.
View latency data
| Documents | p50 · No filter | p99 · No filter |
|---|---|---|
| 10M | 15.5 ms | 17 ms |
| 100M | 31.2 ms | 34.8 ms |
| 1B | 52.9 ms | 61.8 ms |
Latency measured on a production-config cluster with randomly generated documents and queries.
Ship better
search today.
Build on flexible primitives and focus on
what makes your beer taste better.