# State Of The Art Is One Index.

CALDEX was tested on three of the world’s hardest benchmarks, each chosen to break retrieval in a different place: LIMIT where search by meaning runs out, Cohere 1M where memory and speed decide what search costs, and GloVe 1.2M where graph indexes lose their way. One index leads on all three, on 4 cores.

## Unlock more from your vector embeddings.

CALDEX lets you unlock maximum information from your data.

### 01 · LIMIT (Google DeepMind)

[On the Theoretical Limitations of Embedding-Based Retrieval.](https://arxiv.org/abs/2508.21038)

50,000 documents · 1,000 questions, two right answers each · 8 systems from the paper

**Table 1. Recall@2, the share of the true top 2 found.**

| System | Recall@2 (%) |
| --- | ---: |
| **CALDEX**<br>One index · Zerivek | **89.15** |
| BM25<br>Keyword search | 85.7 |
| GTE-ModernColBERT<br>Multi-vector model | 23.1 |
| Promptriever (Llama 3, 8B)<br>Embedding model | 3.2 |
| GritLM 7B<br>Embedding model | 2.4 |
| Gemini Embedding<br>Embedding model | 1.6 |
| E5-Mistral 7B<br>Embedding model | 1.3 |
| Qwen3 Embedding<br>Embedding model | 0.9 |
| Arctic Embed L v2.0<br>Embedding model | 0.4 |

## Make your existing hardware intelligent.

One real-world retrieval benchmark CALDEX dominates.

### 02 · Cohere 1M (VectorDBBench)

One million Cohere embeddings of English Wikipedia.

1,000,000 vectors · 768 dimensions · 10 databases, each on its own machine

**Figure 1. Recall@100, the share of the true top 100 found**

| Database | Recall@100 (%) · more is better |
| --- | ---: |
| **CALDEX**<br>4 cores · 16 GB · VectorDBBench | 98.34% |
| Elastic Cloud<br>8 vCPUs · 60 GB · VectorDBBench | 97.48% |
| Pinecone<br>p2.x8 pod · VectorDBBench | 92.62% |
| zvec<br>16 vCPUs · 64 GB · zvec’s own benchmark | 91.64% |
| turbopuffer<br>Serverless · VectorDBBench | 89.93% |
| pgvector<br>IVFFlat index · 2 vCPUs · 8 GB · VectorDBBench | 88.98% |

**Figure 2. Time to convert raw vectors to searchable knowledge**

| System | Seconds · less is better |
| --- | ---: |
| **CALDEX**<br>4 cores · 16 GB · Single node | 123.6 s |
| zvec<br>16 vCPUs · 64 GB | 147.9 s (1.2× CALDEX’s) |
| Milvus<br>16 vCPUs · 64 GB · HNSW | 581.8 s (4.7× CALDEX’s) |
| Milvus<br>16 vCPUs · 64 GB · SQ8 | 916.3 s (7.4× CALDEX’s) |
| Zilliz Cloud<br>8 CU · performance | 1,016 s (8.2× CALDEX’s) |
| Pinecone<br>p1.x1 pod | 1,124 s (9.1× CALDEX’s) |
| Pinecone<br>s1.x1 pod | 1,129 s (9.1× CALDEX’s) |
| Pinecone<br>p2.x1 pod · 8 nodes | 1,210 s (9.8× CALDEX’s) |
| Pinecone<br>p1.x1 pod · 8 nodes | 1,212 s (9.8× CALDEX’s) |
| Pinecone<br>s1.x1 pod · 2 nodes | 1,451 s (11.7× CALDEX’s) |
| Qdrant Cloud<br>16 vCPUs · 64 GB | 1,500 s (12.1× CALDEX’s) |
| Qdrant Cloud<br>4 vCPUs · 16 GB · 1 node | 1,503 s (12.2× CALDEX’s) |
| Pinecone<br>p2.x1 pod | 1,583 s (12.8× CALDEX’s) |
| Qdrant Cloud<br>4 vCPUs · 16 GB · 5 nodes | 1,648 s (13.3× CALDEX’s) |
| Milvus<br>4 vCPUs · 16 GB · disk | 1,782 s (14.4× CALDEX’s) |
| Pinecone<br>p2.x8 pod · 1 node | 2,126 s (17.2× CALDEX’s) |
| Milvus<br>2 vCPUs · 8 GB · HNSW | 2,380 s (19.3× CALDEX’s) |
| OpenSearch<br>16 vCPUs · 128 GB | 3,572 s (28.9× CALDEX’s) |
| Weaviate Cloud<br>Standard tier | 3,581 s (29.0× CALDEX’s) |
| Weaviate Cloud<br>Business Critical tier | 3,674 s (29.7× CALDEX’s) |
| OpenSearch<br>16 vCPUs · 128 GB · force merge | 4,072 s (32.9× CALDEX’s) |
| Elastic Cloud<br>8 vCPUs · 60 GB | 4,463 s (36.1× CALDEX’s) |
| Milvus<br>2 vCPUs · 8 GB · disk | 6,004 s (48.6× CALDEX’s) |
| Elastic Cloud<br>8 vCPUs · 60 GB · force merge | 6,207 s (50.2× CALDEX’s) |
| Zilliz Cloud | 8,633 s (69.8× CALDEX’s) |
| pgvector<br>2 vCPUs · 8 GB | 10,250 s (82.9× CALDEX’s) |
| Elastic Cloud<br>up to 2.5 vCPUs · 8 GB | 26,410 s (213.7× CALDEX’s) |

**Figure 3. Memory and disk for a searchable index**

| Index | Memory + disk, relative to the raw vectors |
| --- | ---: |
| **CALDEX** | 0.62× (1.6× smaller) |
| HNSW-SQ8 | 1.68× (2.7× CALDEX’s) |
| Elasticsearch | 2.04× (3.3× CALDEX’s) |
| hnswlib | 2.16× (3.5× CALDEX’s) |
| pgvector | ≥ 2.19× (≥ 3.5× CALDEX’s) |
| Weaviate | ≥ 2.21× (≥ 3.6× CALDEX’s) |

**Figure 4. Queries per second against published results**

| Database | Queries/second · more is better |
| --- | ---: |
| **CALDEX**<br>4 cores · 16 GB · VectorDBBench | 2,398 |
| Qdrant Cloud<br>16 vCPUs · 64 GB · VectorDBBench | 1,242 |
| Pinecone<br>p2.x8 pod · VectorDBBench | 1,147 |
| turbopuffer<br>Serverless · VectorDBBench | 798 |
| Weaviate Cloud<br>Business Critical tier · VectorDBBench | 68 |
| pgvector<br>2 vCPUs · 8 GB · VectorDBBench | 11 |

## Less memory. More accuracy.

CALDEX solves the hardest of the graph problems with ease.

### 03 · GloVe 1.2M (ann-benchmarks)

Stanford’s GloVe word vectors, every index built by Zerivek.

1,183,514 vectors · 100 dimensions · 6 other graph indexes

**Table 2. Memory, disk and accuracy comparison, CALDEX vs Existing solutions**

| Index | Accuracy (%) | In memory | On disk | Memory + disk |
| --- | ---: | ---: | ---: | ---: |
| Raw vectors<br>100 dimensions at 32 bits, the vectors alone, no graph | — | — | 1.00× | — |
| *Reach 95% of the true top 10* |  |  |  |  |
| **CALDEX**<br>One index | **96.05** | 0.98× | 0.79× | **1.77×** |
| *Below 95% in this run* |  |  |  |  |
| FAISS HNSW-SQ8<br>8-bit vectors | 94–95 | 1.00× | 2.00× | 3.00× |
| hnswlib<br>32-bit vectors | 92–94 | 1.65× | 1.65× | 3.30× |
| HNSW-Flat<br>32-bit vectors | 89.4 | 1.96× | 1.96× | 3.92× |
| USearch<br>8-bit vectors | 83.69 | 0.69× | 1.69× | 2.38× |
| HNSW-PQ<br>Product-quantised vectors | 78.5 | 0.84× | 1.84× | 2.68× |
| USearch<br>32-bit vectors | 75.7 | 1.65× | 1.65× | 3.30× |

**Table 3. Queries per second on single thread**

| System | Queries/second |
| --- | ---: |
| **CALDEX**<br>4 cores · 16 GB · on a single thread | **1,395** |
| Redis<br>r6i.16xlarge · one CPU · ann-benchmarks | 1,264 |
| Vespa<br>r6i.16xlarge · one CPU · ann-benchmarks | 628 |
| Weaviate<br>r6i.16xlarge · one CPU · ann-benchmarks | 456 |
| Lucene<br>r6i.16xlarge · one CPU · ann-benchmarks | 277 |
| Qdrant<br>r6i.16xlarge · one CPU · ann-benchmarks | 197 |
| pgvector<br>r6i.16xlarge · one CPU · ann-benchmarks | 69 |

## Every result, with the exact place it came from.

CALDEX returns **the file, the exact place and the passage itself**. Go straight from a search result to the evidence in the original source.

*Illustrative results*

**Search:** How much notice must my landlord give before entering my home?

**CALDEX:** Tenancy agreement.pdf › Page 9: The landlord must give **at least 24 hours’ written notice** before entering the property. Opens on page 9, on the sentence that answers.

**Others:** doc_20417 · 0.84. Opens on page 1. The answer is somewhere in 18 pages.

**Search:** How many days do I have to return something I bought online?

**CALDEX:** Terms of sale.pdf › Page 9: You may cancel your order **within 14 days of delivery** without giving a reason. Opens on page 9, on the sentence that answers.

**Others:** doc_73105 · 0.81. Opens on page 1. The answer is somewhere in 14 pages.

**Search:** How much notice do I have to give before I resign?

**CALDEX:** Employment contract.pdf › Page 11: Either party may end this contract by giving **one month’s written notice**. Opens on page 11, on the sentence that answers.

**Others:** doc_11892 · 0.86. Opens on page 1. The answer is somewhere in 22 pages.

**Search:** Can my landlord keep my deposit for normal wear and tear?

**CALDEX:** Tenancy agreement.pdf › Page 13: No deduction may be made from the deposit **for fair wear and tear**. Opens on page 13, on the sentence that answers.

**Others:** doc_20417 · 0.79. Opens on page 1. The answer is somewhere in 18 pages.

**Search:** What is the late fee if I pay my rent after the due date?

**CALDEX:** Tenancy agreement.pdf › Page 10: A late fee of **5% of the monthly rent** applies from the sixth day after the due date. Opens on page 10, on the sentence that answers.

**Others:** doc_20417 · 0.82. Opens on page 1. The answer is somewhere in 18 pages.

**Search:** How long is the warranty on my laptop?

**CALDEX:** Warranty terms.pdf › Page 9: This product is covered against defects **for 24 months from the date of purchase**. Opens on page 9, on the sentence that answers.

**Others:** doc_56230 · 0.88. Opens on page 1. The answer is somewhere in 20 pages.

**Search:** Can I end my lease early if I move for work?

**CALDEX:** Tenancy agreement.pdf › Page 15: The tenant may end this tenancy early by giving **60 days’ notice and one month’s rent**. Opens on page 15, on the sentence that answers.

**Others:** doc_20417 · 0.77. Opens on page 1. The answer is somewhere in 18 pages.

**Search:** How long do I have to report a wrong charge on my credit card?

**CALDEX:** Cardholder agreement.pdf › Page 21: You must tell us about a billing error **within 60 days of the statement date**. Opens on page 21, on the sentence that answers.

**Others:** doc_90344 · 0.83. Opens on page 1. The answer is somewhere in 32 pages.

**Search:** How long does my non-compete last after I leave?

**CALDEX:** Employment contract.pdf › Page 17: For **12 months after your employment ends**, you may not work for a direct competitor. Opens on page 17, on the sentence that answers.

**Others:** doc_11892 · 0.80. Opens on page 1. The answer is somewhere in 22 pages.

**Search:** Does my car insurance cover a hire car while mine is repaired?

**CALDEX:** Motor insurance policy.pdf › Page 19: We will pay for a hire car **for up to 21 days** while your car is repaired after a claim. Opens on page 19, on the sentence that answers.

**Others:** doc_38761 · 0.85. Opens on page 1. The answer is somewhere in 44 pages.

**Search:** How many days of paid sick leave do I get a year?

**CALDEX:** Employee handbook.pdf › Page 15: Full-time employees receive **ten days of paid sick leave** each calendar year. Opens on page 15, on the sentence that answers.

**Others:** doc_48213 · 0.84. Opens on page 1. The answer is somewhere in 48 pages.

**Search:** How much unused leave can I carry over?

**CALDEX:** Employee handbook.pdf › Page 12: Unused leave carries over for **up to five days** into the next year. Opens on page 12, on the sentence that answers.

**Others:** doc_48213 · 0.87. Opens on page 1. The answer is somewhere in 48 pages.

Source: https://www.zerivek.com/docs/benchmarks