Voyage AI Embeddings: Complete Guide — Architecture, MTEB Benchmarks, API, Pricing & RAG Optimization 2026

Model Overview

Voyage AI Embeddings (voyage-2, voyage-3, and voyage-large-2) is a family of proprietary text embedding models developed by Voyage AI, a company founded in 2023 by Stanford researchers and acquired by Anthropic in 2025. The flagship voyage-2 model produces 1,024-dimensional vectors and achieves a 63.8 MTEB average score, with voyage-3 improving to 64.5. Voyage AI belongs to the text embedding model category and solves the problem of measuring semantic similarity between text with models specifically optimized for retrieval-augmented generation (RAG). Voyage AI is designed for developers building production RAG pipelines who want embeddings co-optimized with retrieval quality rather than general semantic similarity. Its key differentiators are RAG-specific training (optimized for retrieval rather than general similarity), long context support (up to 32,000 tokens in voyage-3), and native integration with Anthropic's Claude models. In 2026, Voyage AI is the recommended embedding provider for Anthropic-stack RAG pipelines and is deeply integrated into Claude's retrieval features. The Anthropic acquisition has accelerated its adoption, making it a top-tier commercial embedding option alongside OpenAI and Cohere.

Architecture & Technical Deep Dive

Voyage AI uses a transformer-based encoder architecture trained with a contrastive learning objective specifically optimized for retrieval-augmented generation. The models are proprietary and designed to produce embeddings that maximize retrieval quality in RAG pipelines rather than general semantic similarity.

Core Architecture

Voyage AI has not fully disclosed the architecture, but the models use a transformer encoder producing 1,024-dimensional vectors. The encoder processes text up to 4,000 tokens (voyage-2) or 32,000 tokens (voyage-3) and produces a single dense vector per input. The 32K token context in voyage-3 is a significant differentiator — most embedding models max out at 512-8K tokens, making voyage-3 ideal for embedding long documents without chunking. The model uses attention pooling over token representations to generate the final embedding.

RAG-Optimized Training

Voyage AI's key innovation is RAG-specific training. While most embedding models are trained for general semantic similarity across MTEB tasks, Voyage AI optimizes for retrieval quality — the ability to find relevant documents for a given query in a RAG pipeline. Training uses contrastive learning with a focus on query-document relevance, hard-negative mining, and domain-specific retrieval data. This means Voyage AI may score slightly lower on some MTEB subtasks (STS, clustering) but higher on retrieval-specific benchmarks. The models are co-optimized with reranking models from Voyage AI for end-to-end RAG quality.

Long Context Support

Voyage-3 supports up to 32,000 tokens per input — significantly longer than OpenAI (8,191), Cohere (512), BGE (512), and Jina (8,192). This enables embedding entire long documents (research papers, legal contracts, technical documentation) in a single vector without chunking. Long context reduces information loss from chunking, improves retrieval of information that spans chunk boundaries, and simplifies pipeline architecture. For documents under 4K tokens, voyage-2 and voyage-3 perform similarly; for longer documents, voyage-3 is significantly better.

Training Methodology

Voyage AI uses a dual-encoder contrastive learning approach. The model is trained on query-document pairs from web search, enterprise documents, and domain-specific corpora. Training emphasizes hard-negative mining — the model learns to distinguish relevant documents from superficially similar but irrelevant ones. Voyage AI has not disclosed the exact training corpus size or compute budget. The company has published research papers on retrieval optimization and RAG quality, providing more transparency than some commercial competitors. The voyage-3 update improved long-context handling and retrieval quality over voyage-2.

Inference Requirements

Voyage AI is API-only — no public local deployment. Inference is handled on Voyage AI's cloud or via AWS Bedrock. Latency: 200-500ms per request for voyage-2, 300-800ms for voyage-3 (longer context). Throughput: up to 2,000 RPM on production tier. Batch API processes large corpora with cost reduction. For enterprise customers, Voyage AI offers custom deployment options. No VRAM or hardware requirements for API users. The Anthropic acquisition may bring tighter integration with Anthropic's infrastructure.

MTEB Benchmark Performance & Scores

Scores based on publicly available data as of July 2026. Independent verification recommended.

MTEB Benchmark Comparison

Speed & Throughput

Latency: 200-500ms per request for voyage-2 (up to 4K tokens), 300-800ms for voyage-3 (up to 32K tokens). Batch API available for bulk embedding generation with cost reduction. Throughput: 2,000 RPM (production tier). For high-volume production, use Batch API for initial corpus embedding. For real-time RAG, use the standard API with caching for repeated queries. The 32K context in voyage-3 reduces chunking overhead for long documents.

Speed & Latency

Latency: 200-500ms per request for voyage-2 (up to 4K tokens), 300-800ms for voyage-3 (up to 32K tokens). Batch API available for bulk embedding generation with cost reduction. Throughput: 2,000 RPM (production tier). For high-volume production, use Batch API for initial corpus embedding. For real-time RAG, use the standard API with caching for repeated queries. The 32K context in voyage-3 reduces chunking overhead for long documents.

API Access, Pricing & Integration Guide

Looking for voyage-2 API pricing in 2026? Below is the complete pricing table, code examples, and integration guide.

API Pricing Table (as of July 2026)

Free Tier & Trial Access

Voyage AI offers a free tier: 50M tokens free for new accounts (200M tokens for Anthropic Claude customers). This is one of the most generous free tiers among commercial embedding providers. Batch API offers 50% discount for non-urgent embedding generation. Enterprise pricing available for high-volume customers.

API Quick Start

# Install SDK
pip install voyageai

import voyageai
import numpy as np

vo = voyageai.Client(api_key="your-api-key")

# Embed documents for indexing
doc_result = vo.embed(
    texts=["Machine learning is fascinating",
           "AI transforms industries",
           "I love pizza"],
    model="voyage-3",
    input_type="document"  # Optimized for indexing
)
doc_embeddings = doc_result.embeddings

# Embed a query
query_result = vo.embed(
    texts=["What is AI?"],
    model="voyage-3",
    input_type="query"  # Optimized for retrieval
)
query_embedding = query_result.embeddings[0]

# Compute cosine similarity
sims = np.dot(doc_embeddings, query_embedding) / (
    np.linalg.norm(doc_embeddings, axis=1) * np.linalg.norm(query_embedding))
print(f"Similarities: {sims}")

Supported API Features

Compatible Platforms & Integrations

Fine-Tuning, RAG & Advanced Use

Fine-Tuning Availability

Voyage AI does not offer public fine-tuning via API. Enterprise customers can work with Voyage AI for custom-trained embeddings on domain-specific data. For fine-tunable open-source embeddings, consider BGE-large, E5-large, or Sentence Transformers with LoRA adaptation. Voyage AI's RAG-optimized training and input type specification partially compensate — the model is already optimized for retrieval without fine-tuning.

Fine-Tuning Requirements

N/A — fine-tuning not available via public API. Enterprise customers can contact Voyage AI or Anthropic for custom embedding training with domain-specific corpora. For self-service fine-tuning, use open-source alternatives: BGE-large (LoRA fine-tuning), E5-large (contrastive fine-tuning), or Sentence Transformers (highly customizable). Recommended training data: 50,000+ domain-specific query-document pairs.

Fine-Tuning Use Cases

• RAG pipeline embedding layer — embed documents and queries with RAG-optimized embeddings for retrieval-augmented generation with Claude
• Long document retrieval — use voyage-3's 32K context to embed entire documents without chunking, reducing information loss
• Anthropic-stack RAG — pair Voyage AI embeddings with Claude for end-to-end Anthropic ecosystem RAG with native integration
• Enterprise knowledge base search — power internal search with Voyage AI embeddings and Claude for grounded, citation-backed answers
• Legal and technical document search — use long context to embed entire contracts, papers, and technical docs in a single vector

RAG Integration Guide

Voyage AI is specifically optimized for RAG. Recommended architecture: Documents → (optional chunking for >32K docs) → voyage-3 (input_type="document") → Vector DB (Pinecone, Weaviate, Qdrant) → Query voyage-3 (input_type="query") → Retriever → Claude (Anthropic). The input_type specification is important — use "document" for indexing and "query" for retrieval. For long documents, voyage-3's 32K context enables embedding entire documents without chunking. Pair with Voyage AI's reranker for end-to-end RAG optimization. Batch API for initial corpus embedding; standard API for real-time queries.

Prompt Engineering Tips

• Always use input_type="document" when embedding documents for indexing
• Always use input_type="query" when embedding user queries for retrieval
• Use voyage-3 for long documents (up to 32K tokens) to avoid chunking information loss
• Use voyage-2 for documents under 4K tokens — faster and similar quality
• Pair with Voyage AI reranker for end-to-end RAG optimization
• Batch API for initial corpus embedding — 50% cost reduction
• Use the 50M free tokens for initial prototyping and testing

Use Cases, Strengths & Limitations

Top 10 Real-World Use Cases

Strengths

• ✓ RAG-Optimized Training — specifically trained for retrieval quality, not just general semantic similarity
• ✓ Long Context (32K tokens) — voyage-3 handles entire documents without chunking, reducing information loss
• ✓ Native Anthropic Integration — tightest integration with Claude among embedding providers
• ✓ Top-Tier MTEB Performance — 64.5 MTEB (voyage-3) ranks among top commercial embedding models
• ✓ Generous Free Tier — 50M tokens free (200M for Claude customers), most generous among commercial providers
• ✓ Input Type Specification — asymmetric embeddings optimized for retrieval
• ✓ Batch API — 50% cost reduction for bulk embedding generation
• ✓ AWS Bedrock Deployment — available on AWS Bedrock for enterprise compliance

Limitations & Weaknesses

• ✗ No Public Fine-Tuning — cannot customize the model for domain-specific vocabulary via API
• ✗ No Local Deployment — proprietary model; data must be sent to Voyage AI or AWS Bedrock
• ✗ Limited Multilingual — English-focused; multilingual support expanding but not as strong as Cohere
• ✗ Higher Latency — 200-800ms per request, slower than OpenAI (100-300ms) due to longer context
• ✗ Smaller Ecosystem — fewer third-party integrations than OpenAI, though growing with Anthropic acquisition
• ✗ Fixed 1,024 Dimensions — no Matryoshka truncation like OpenAI text-embedding-3
• ✗ No Image/Multimodal Embeddings — text-only; for multimodal, use CLIP-based models

Who Should Use This Model

Alternatives, Comparisons & Verdict

Top Alternatives

Detailed Comparison

Voyage AI vs OpenAI text-embedding-3-large: Both score nearly identically on MTEB (64.5 vs 64.6). Voyage AI offers 32K context (vs OpenAI's 8K) and RAG-optimized training; OpenAI offers flexible Matryoshka dimensions and larger ecosystem. Voyage AI has native Claude integration; OpenAI has native GPT-4o integration. Voyage AI costs $0.12/1M; OpenAI costs $0.13/1M. Choose Voyage AI for long context and Claude integration, OpenAI for flexible dimensions and ecosystem. Voyage AI vs Cohere Embed v3: Both score identically on MTEB (64.5). Voyage AI offers 32K context and RAG-optimized training; Cohere offers 100+ language multilingual support and input type specification. Cohere costs slightly less ($0.10 vs $0.12/1M). Choose Voyage AI for long context and English RAG, Cohere for multilingual search.

Our Verdict

Internal Links

Frequently Asked Questions