## Monday, February 9, 2026

# Voyage 4: Evaluation Notes

Umida Muratbekova  
RAGEmbeddings  
Voyage AI recently released **[Voyage 4](/content/embeddings/voyage-4/index.html)**. We spent some time running it through our evaluation pipeline to see how it shifts the needle for RAG compared to [Voyage 3 Large](/content/embeddings/voyage-3-large/index.html) and current frontier embedding models.

Here is what we found.

## The Evaluation

We tested 14 models across 6 datasets, covering financial Q&A, web search, scientific claims, and entity retrieval. We used two primary lenses:

1. Traditional metrics: NDCG@10.
2. LLM-as-a-judge: Pairwise comparisons aggregated into ELO ratings. We had GPT-5 compare the Top-5 results from model pairs to determine which set provided better context for the query.

## **Key Observations**

[Voyage 4](/content/embeddings/voyage-4/index.html) is currently the strongest general-purpose model we've tested.

- Win Rate: 61.7% across 780 head-to-head comparisons.
- Elo: 1606 (the highest in our current pool).
- Precision: 0.859 NDCG@10.

The most interesting signal is the Top-K density. [Voyage 4](/content/embeddings/voyage-4/index.html)'s advantage over its predecessors grows by 2.3× when you evaluate only the Top-5 results instead of the Top-10. It's significantly more "top-heavy" with relevant information.

## Domain Performance: Factual vs. Reasoning

The model isn't a linear upgrade across every domain. It has a very specific signature.

- **Entity-Rich & Business Documents**: The model is strongest here. It performed exceptionally well on DBPedia and Business Reports, making it ideal for finding specific data points in professional documents.
- **The SciFact Ceiling**: This was the only outlier. [Voyage 4](/content/embeddings/voyage-4/index.html) dropped to 1465 Elo with a 90% tie rate. It seems to hit a wall on tasks requiring subtle scientific reasoning or evidence-contradiction analysis. It behaved like a high-end semantic matcher, but not a "reasoner."

## The OpenAI Parity

We found a perfect stalemate between [Voyage 4](/content/embeddings/voyage-4/index.html) and [OpenAI text-embedding-3-large](/content/embeddings/openai-text-embedding-3-large/index.html).

- **Record**: 28 wins – 28 losses – 4 ties.  
Neither model could establish dominance over the other across 60 queries. They currently sit in an elite tier together, roughly 60+ Elo points above the rest of the field.

## Retrieval Quality in RAG

Most RAG systems are optimized for Top-3 or Top-5 chunks to reduce noise and context costs. This is where [Voyage 4](/content/embeddings/voyage-4/index.html) makes the most sense.

We looked at a specific query regarding post-2000 gold economics:

- [Voyage 4](/content/embeddings/voyage-4/index.html): Returned 5/5 chunks directly addressing gold price drivers (QE3, real interest rates, financial crisis).
- [Voyage 3 Large](/content/embeddings/voyage-3-large/index.html): Returned 0/5 relevant chunks. It surfaced "finance-adjacent" noise - Hyundai cruise control pricing and IRS tax procedures.

The semantic overlap in [Voyage 3 Large](/content/embeddings/voyage-3-large/index.html) was too broad. [Voyage 4](/content/embeddings/voyage-4/index.html) surfaced answerable context where the predecessor only found topical keywords.

## Conclusion

[Voyage 4](/content/embeddings/voyage-4/index.html) is currently the **strongest general-purpose model** on our [leaderboard](/content/embeddings/index.html). If your RAG system relies on factual retrieval and you are optimizing for a small context window (Top-5), this should be the new default.

Full dataset breakdowns and rankings are at [the embeddings leaderboard](/content/embeddings/index.html).