README

Documents

Quick Recommendation

Performance Rankings (MTEB Benchmark)

Top Commercial Models

Top Open-Source Models

Detailed Pricing (November 2025)

Commercial APIs

Free/Open-Source Options

Key Findings

Voyage AI: Current Leader

• voyage-3-large ranks #1 across 8 domains (law, finance, code, etc.)
• Outperforms OpenAI-v3-large by 9.74% average
• Outperforms Cohere-v3-English by 20.71% average
• 32K context window (vs OpenAI’s 8K, Cohere’s 512)
• Supports Matryoshka learning: use smaller dimensions (256, 512, 1024) with minimal quality loss
• At 1/24 the storage cost (int8 512 dims), still beats OpenAI by 8.56%

OpenAI: Best Value for Quality

• text-embedding-3-small at $0.02/1M tokens is exceptional value
• Batch API cuts cost in half ($0.01/1M tokens)
• Well-documented, reliable, wide ecosystem support
• Not the absolute best quality, but “good enough” for most use cases

Open-Source: Nomic Embed v2

• First general-purpose Mixture of Experts (MoE) embedding model
• Outperforms models 2x its size on multilingual benchmarks
• Fully open: weights, training data, and code all available
• Supports ~100 languages
• Flexible dimensions (768 to 256) via Matryoshka representation
• Run locally via Ollama: ollama pull nomic-embed-text

Cohere: Best for Multilingual/Noisy Data

• Handles 100+ languages with consistent quality
• Robust to noisy, real-world data (typos, OCR errors)
• Higher price point but reliable for enterprise

Warning: MTEB Benchmark Gaming

Many open-source models appear to be fine-tuned specifically on MTEB benchmarks, producing inflated scores that don’t reflect real-world performance. Best practice is downstream evaluation - measure actual retrieval accuracy in your application.

Detailed Performance: voyage-3.5-lite vs OpenAI Models

MTEB Benchmark Scores

Performance Analysis

voyage-3.5-lite advantages over OpenAI:

• Outperforms text-embedding-3-large by 6.34% on average across domains
• Outperforms text-embedding-3-small by ~7.58% at the same price point
• Achieves retrieval quality within 0.3% of Cohere-v4 at 1/6 the cost
• 83% reduction in vector database costs vs OpenAI 3-large (using int8 2048-dim vs float 3072-dim)

Storage Efficiency Comparison

Real-World Migration Case Study (MyClone)

MyClone migrated from OpenAI text-embedding-3-small to voyage-3.5-lite:

• 3× storage savings (1536-dim → 512-dim via Matryoshka)
• 2× faster retrieval due to smaller vectors
• 15-20% latency reduction in voice applications
• No quality loss - Matryoshka training preserves semantic signal in lower dimensions

Why voyage-3.5-lite Wins

1. Same price, better quality: At $0.02/1M tokens, matches OpenAI 3-small pricing but delivers 7-8% better accuracy
2. Flexible dimensions: Matryoshka learning allows 256/512/1024/2048 dimensions with minimal quality loss
3. Larger context: 32K tokens vs OpenAI’s 8K (4x improvement)
4. Storage optimization: int8 quantization + smaller dimensions = 83% cost reduction vs OpenAI 3-large
5. Free tier: 200M tokens free vs no free tier for OpenAI

When to Still Use OpenAI

• Batch API pricing: $0.01/1M tokens is 50% cheaper than voyage-3.5-lite (only advantage on cost)
• Ecosystem integration: Better SDK/library support in some frameworks (LangChain, LlamaIndex)
• Existing infrastructure: Already using OpenAI APIs extensively
• Vendor consolidation: Want all AI services from one provider
• ada-002 migration: Direct drop-in replacement with 3-small (same 1536 dims)

Note: At standard pricing ($0.02/1M), there’s no cost advantage - voyage-3.5-lite wins on both price AND quality.

Cost Analysis: 1 Million Documents

Assuming 1,000 tokens average per document (1B tokens total):

Note: Voyage AI offers 200M free tokens for new accounts, reducing initial costs.

Recommendations by Scenario

For @research/graph (Our Use Case)

Option 1: Voyage voyage-3.5-lite (Recommended)

• $0.02/1M tokens - same price as OpenAI 3-small
• 7-8% better accuracy than OpenAI at same cost
• 32K context window handles full documents
• 200M free tokens for new accounts
• Flexible dimensions (256-2048) for storage optimization

Option 2: OpenAI text-embedding-3-small (If Already Using OpenAI)

• Batch API: $0.01/1M tokens (50% cheaper than voyage-3.5-lite)
• Well-supported in TypeScript/Node.js ecosystem
• Easy API integration, wide library support
• Only choose over Voyage if: using batch API OR already on OpenAI infrastructure

Option 3: Nomic Embed via Ollama (Cost-Free)

• Zero API costs
• Privacy: data never leaves your machine
• Slightly more complex setup
• May need GPU for speed

Option 4: Voyage voyage-3-large (Best Quality)

• If semantic search quality is critical
• 32K context handles full documents
• Higher cost ($0.18/1M) justified for production RAG

For Production RAG Systems

1. Best overall value: Voyage voyage-3.5-lite ($0.02/1M, best accuracy at this price)
2. Absolute cheapest API: OpenAI batch API ($0.01/1M) - 50% cheaper but 7% less accurate
3. Cost-free: Nomic local via Ollama or HuggingFace TEI
4. Quality-critical: Voyage voyage-3-large ($0.18/1M, top MTEB scores)
5. Multilingual enterprise: Cohere embed-v3 (~$0.40/1M, 100+ languages)
6. Air-gapped/private: Nomic, BGE, or E5 via Hugging Face TEI

Bottom line: Don’t use OpenAI standard pricing - either use batch API (cheaper) or switch to Voyage (better).

Running Local Embeddings

Ollama (Simplest)

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull embedding model
ollama pull nomic-embed-text

# Generate embeddings
curl http://localhost:11434/api/embeddings \
  -d '{"model": "nomic-embed-text", "prompt": "Hello world"}'

Hugging Face Text Embeddings Inference (Production)

docker run --gpus all -p 8080:80 \
  ghcr.io/huggingface/text-embeddings-inference:latest \
  --model-id BAAI/bge-base-en-v1.5

Voyage AI: Query vs Document Input Types

Voyage AI’s embedding API includes an input_type parameter that significantly impacts search quality for asymmetric retrieval tasks.

The input_type Parameter

Why It Matters

For asymmetric retrieval (short queries matching longer documents):

• Documents should be embedded with input_type: "document"
• Search queries should be embedded with input_type: "query"

Using "document" for both produces embeddings in different semantic spaces, causing poor search results.

Impact on Search Quality

Before (using “document” for queries):

Search: "bun link"
1. Pluribus (poker AI) - 84.21%  ❌ Irrelevant
2. Libratus (poker AI) - 82.57%  ❌ Irrelevant
3. bun link - 82.07%

After (using “query” for queries):

Search: "bun link"
1. bun link - 69.90%  ✅ Correct
2. bun-link.md - 65.06%  ✅ Correct
3. Bun Package Manager - 57.25%  ✅ Correct

The absolute similarity scores are lower, but the ranking is correct - which is what matters for search.

Compatibility

Voyage states that “embeddings generated with and without the input_type argument are compatible” - they can be compared via cosine similarity even when generated with different input types. However, using the appropriate type for each use case produces better retrieval results.

Implementation

// For storing documents
const docEmbedding = await voyageAPI.embed(text, { input_type: "document" });

// For search queries
const queryEmbedding = await voyageAPI.embed(query, { input_type: "query" });

Sources

• Voyage AI: Text Embeddings Documentation
• Voyage AI: voyage-3.5 and voyage-3.5-lite Announcement
• Voyage AI: voyage-3-large Announcement
• Voyage AI: Pricing
• OpenAI: New Embedding Models and API Updates
• OpenAI Pricing
• MongoDB: Introducing voyage-3.5 and voyage-3.5-lite
• MyClone: How We Cut RAG Latency by 50% with Voyage 3.5 Lite
• Elephas: 13 Best Embedding Models in 2025
• Pinecone: Choosing an Embedding Model
• Document360: Text Embedding Models Compared
• Nomic AI: Introducing Nomic Embed
• BentoML: Guide to Open-Source Embedding Models
• Supermemory: Best Open-Source Embedding Models Benchmarked
• Hugging Face: Text Embeddings Inference

Last updated: 2025-12-07