Embedding and rerank - Alibaba Cloud Model Studio - Alibaba Cloud Documentation Center

Migrate from closed-source models

Map your current OpenAI, Cohere, or Voyage embedding/reranking model to an equivalent Bailian model.

	Closed-source examples	Bailian recommendation
Text embedding	OpenAI text-embedding-3-large, Voyage-3-large, Cohere embed-v4	`text-embedding-v4`
Multimodal embedding	Cohere embed-v4, Voyage multimodal	`qwen3-vl-embedding`, `tongyi-embedding-vision-plus`
Reranking	Cohere Rerank 3.5, Voyage rerank-2.5	`qwen3-rerank`, `gte-rerank-v2`

Text embedding

For plain text search, RAG, or clustering, use text-embedding-v4. If you need to migrate existing v3 indexes, use text-embedding-v3 for dimension compatibility.

How to choose a dimension

For large-scale search with limited storage: choose 256 or 512 dimensions.
For general-purpose use cases: choose 1024 dimensions (the default, offering a good balance).
For high retrieval accuracy: choose 1536 or 2048 dimensions.

Multimodal embedding

For cross-modal retrieval, such as text-to-image or text-to-video search, choose between fused and independent vectors.

Fused vs. independent vectors

Fused vector: Fuses text and images into a single vector for mixed text-and-image retrieval. Use qwen3-vl-embedding.
Independent vector: Generates a separate vector for each modality. Suitable for cross-modal searches (text-to-image, image-to-image). Use tongyi-embedding-vision-plus.

Working with text-only data?

Use text-embedding-v4 — it is faster, more cost-effective, and offers more dimension options. Multimodal embedding is for cross-modal retrieval such as text-to-image and text-to-video search.

Rerank

After embedding-based retrieval, use a rerank model to reorder the top-N results with cross-attention for higher accuracy.

Plain text rerank: Uses qwen3-rerank, with support for over 100 languages and up to 500 documents.
Multimodal reranking: Uses qwen3-vl-rerank to rerank a mix of text, images, and videos.

All models

Model ID	Type	Dimension	Max tokens	Use case
`text-embedding-v4`	Text embedding	64–2048 (default: 1024)	8,192	Text search, RAG, clustering
`text-embedding-v3`	Text embedding	512–1024 (default: 1024)	8,192	Migrating existing v3 indexes
`tongyi-embedding-vision-plus`	Multimodal embedding	64–1152 (default: 1152)	1,024	Cross-modal search (independent vector only)
`tongyi-embedding-vision-flash`	Multimodal embedding	64–768 (default: 768)	1,024	Cost-sensitive cross-modal search (independent vector only)
`qwen3-rerank`	Rerank	-	4,000 per item	Reranking text search results, RAG
`qwen3-vl-embedding`	Multimodal embedding	256–2560 (default: 2560)	32,000	Mixed text-and-image retrieval (fused vector + independent vector)
`qwen3-vl-rerank`	Rerank	-	8,000 per item	Reranking multimodal search results