Azure Text Embedding Model - Overview

Costa Rica

brown9804

Last updated: 2025-08-27

---

List of References (Click to expand)

• voyage-3-large: the new state-of-the-art general-purpose embedding model
• text-embedding-3-large
• Embedding Leaderboard - reports
• The Legacy MTEB Leaderboard repository - explanation and relevant links
• Massive Text Embedding Benchmark - open source how it was created
• Model leaderboards in Azure AI Foundry portal (preview)
• Find the best model by comparing model performance across various criteria
• Chunk large documents for vector search solutions in Azure AI Search

Table of Contents (Click to expand)

• Recommendations for Alternative & Upcoming Embedding Models in Azure
  • Azure OpenAI
  • Azure AI Foundry
  • Azure Marketplace
• How Azure AI Search Can Help Optimize Embedding Performance

Recommendations for Alternative & Upcoming Embedding Models in Azure

Current Options:

Model	Platform	Strengths	Limitations	Best Use Cases
voyage-3	Azure Marketplace	Balanced performance, tuned for enterprise domains, cost-efficient	No native integration with Azure AI Foundry, limited multilingual support	Internal search, domain-specific retrieval, enterprise RAG
voyage-3-lite	Azure Marketplace	Lightweight, fast, low latency and compute cost	Slightly reduced semantic precision, less robust for long-form text	Mobile apps, real-time classification, edge deployments
voyage-finance-2	Azure Marketplace	Specialized for financial language and jargon, tuned for structured documents	Poor generalization outside finance, narrow scope	Financial document analysis, compliance, financial RAG
voyage-law-2	Azure Marketplace	Legal-specific, optimized for contracts and statutes	Limited cross-domain utility, narrow vocabulary	Legal search, contract parsing, regulatory compliance
voyage-multilingual-2	Azure Marketplace	Multilingual support, tuned for global corpora	Slightly lower precision in English, not ideal for domain-specific tasks	Multilingual semantic search, global content indexing
text-embedding-ada-002	Azure OpenAI / AI Foundry	Legacy model, fast and inexpensive, widely supported	Lower semantic quality, outdated compared to newer models	Lightweight search, prototyping, low-cost RAG
text-embedding-3-small	Azure OpenAI / AI Foundry	Fast, cost-effective, configurable, good for short texts	Lower accuracy on nuanced or complex queries	Chat summarization, real-time applications, indexing
text-embedding-3-large	Azure OpenAI / AI Foundry	High semantic fidelity, multilingual, robust across domains	Higher latency and cost, batch quirks	Semantic search, multilingual corpora, recommendation systems
embed-v-4-0	Azure AI Foundry	High-performance, optimized for retrieval and RAG, scalable	Newer model with limited public benchmarks	Advanced RAG pipelines, enterprise search, hybrid retrieval
Cohere-embed-v3-multilingual	Azure AI Foundry	Strong multilingual alignment, semantic robustness	Slightly slower, less tuned for English-only tasks	Multilingual indexing, global search, translation-aware retrieval
Cohere-embed-v3-english	Azure AI Foundry	High precision for English, optimized for semantic tasks	Not suitable for multilingual content	English-centric semantic search, document clustering

Note

Upcoming Models to Watch:

• Voyage-3-Large: Expected to outperform OpenAI v3-large with flexible dimensions and quantization. voyage-3-large: the new state-of-the-art general-purpose embedding model

Azure OpenAI

• text-embedding-ada-002
• text-embedding-3-large
• text-embedding-3-small

1. Go to your Azure OpenAI Platform:

   

2. Under Model catalog, filter by Inference task > Embeddings:

   

Azure AI Foundry

• embed-v-4-0
• Cohere-embed-v3-multilingual
• Cohere-embed-v3-english
• text-embedding-ada-002
• text-embedding-3-large
• text-embedding-3-small

1. Go to your Azure AI Foundry Platform:

   

2. Under Model catalog, filter by Inference task > Embeddings:

   

Azure Marketplace

Azure Marketplace:

• voyage-3 Embedding Model
• voyage-3-lite Embedding Model
• voyage-finance-2 Embedding Model
• voyage-law-2 Embedding Model
• voyage-multilingual-2 Embedding Model

How Azure AI Search Can Help Optimize Embedding Performance

Azure AI Search doesn’t generate embeddings from third-party models, but it amplifies their value through advanced indexing, retrieval, and hybrid search capabilities.

Tip

Use Azure AI Foundry to experiment with multiple embedding models and benchmark their performance before committing to production.

Vector Search Capabilities (Click to expand)

Azure AI Search supports robust vector search features that allow you to fully leverage embeddings generated externally (e.g., from OpenAI, Voyage, Cohere):

• Accepts custom embeddings from any model, making it agnostic and flexible across providers.
• Supports semantic similarity search using metrics like cosine similarity and dot product, enabling nuanced matching beyond keyword overlap.
• Handles large-scale indexing and low-latency retrieval, ideal for enterprise-grade applications with millions of documents.
• Integrates with hybrid search pipelines, combining vector and keyword search to improve both precision and recall.

Example Use Cases:

• Intelligent document retrieval for legal or financial archives.
• FAQ matching and chatbot grounding using semantic similarity.
• Product recommendation systems based on user intent embeddings.

Performance Optimization Tips (Click to expand)

To maximize efficiency and relevance in embedding-based search, consider these strategies:

1. Choose embedding size wisely: Smaller embeddings (e.g., 512–1024) reduce latency and storage costs, ideal for mobile or real-time apps.
2. Batch embedding generation: Pre-process documents in bulk to reduce API calls and improve throughput.
3. Use domain-specific models: Models like voyage-finance-2 or voyage-law-2 yield better semantic relevance in specialized contexts.
4. Monitor vector DB costs: Larger embeddings increase storage and query costs, balance precision with efficiency.
5. Leverage hybrid search: Combine keyword and vector search to handle both exact and fuzzy matches, especially in noisy datasets.
6. Normalize and deduplicate embeddings: Ensure consistent vector quality and avoid redundant indexing.

Integration Strategy (Click to expand)

Azure AI Search is designed to integrate seamlessly with external embedding pipelines. Here's how to build a modular and scalable setup:

• Generate embeddings externally using models from Azure OpenAI (text-embedding-3-large, text-embedding-3-small), Voyage AI (voyage-3, voyage-multilingual-2), or Cohere (embed-v3 series).
• Store embeddings in Azure AI Search vector fields, which are optimized for fast similarity search and scalable indexing.
• Query using embedded vectors to perform semantic matching, enabling intelligent document retrieval, contextual search, and RAG workflows.
• Automate updates: Use Azure Functions or Logic Apps to refresh embeddings when documents change, keeping your index up-to-date.

Example Workflow:

1. Use text-embedding-3-large to embed support tickets.
2. Store vectors in Azure AI Search.
3. Query with user questions to retrieve semantically similar tickets.
4. Combine with keyword filters for precision.

Refresh Date: 2025-08-27