Embedding - Vectorization Model Selection

Convert text, images, and other data into vectors - the foundation of all vector retrieval applications.

Quick Selection

By User Type

User Type	Recommended	Reason
Beginners (don't want to manage infrastructure)	OpenAI / Cohere API	Sign up and use, no GPU needed
Advanced users (seeking best results)	BGE-M3 / Voyage AI	Open source control / best for specialized domains
Chinese scenarios	BGE-M3 or Qwen3-Embedding	Best Chinese performance
Image scenarios	CLIP / Chinese-CLIP	Cross-modal image-text

By Scenario

Scenario	Recommended Model	Notes
General text search	OpenAI text-embedding-3-small	Best value
High-precision RAG	Voyage-3-large / Cohere embed-v4	MTEB top performers
Chinese knowledge base	BGE-M3 / BGE-large-zh	Chinese optimized
Multilingual	BGE-M3 / multilingual-e5	100+ languages
Code search	Voyage-code-3	Code specialized
Legal/Finance	Voyage-finance/law	Domain specialized
Image search	CLIP-ViT-L-14	General image-text
Chinese image-text	Chinese-CLIP	Chinese image-text

---

Model Overview

API Models (Ready to Use)

Suitable for: Beginners, no GPU management, quick validation

Model	Provider	Dimensions	Context	Price ($/1M tokens)	MTEB	Features
text-embedding-3-large	OpenAI	3072	8K	$0.13	64.6	Best general, mature ecosystem
text-embedding-3-small	OpenAI	1536	8K	$0.02	62.3	Best value
embed-v4	Cohere	1024	512	$0.10	65.2	Multilingual, compressible
voyage-3-large	Voyage AI	1024	32K	$0.06	65.8	Long text, domain specialized
voyage-3-lite	Voyage AI	512	32K	$0.02	61.2	Low cost, low latency

Open Source Models (Local Deployment)

Suitable for: Advanced users, data privacy requirements, seeking best performance

Model	Source	Dimensions	Context	Size	MTEB	Features
BGE-M3	BAAI	1024	8K	2.2GB	63.0	Multi-lingual/functional/granular, best Chinese
BGE-large-zh-v1.5	BAAI	1024	512	1.3GB	-	Pure Chinese, lightweight
Qwen3-Embedding-8B	Alibaba	4096	32K	16GB	66.2	Latest, best performance
Qwen3-Embedding-0.6B	Alibaba	1024	32K	1.2GB	58.5	Lightweight, surpasses BGE-M3 at same size
jina-embeddings-v3	Jina AI	1024	8K	1.2GB	62.8	Multi-task, adjustable dimensions
nomic-embed-text	Nomic	768	8K	548MB	59.3	Open source free, local first choice
all-MiniLM-L6-v2	SBERT	384	256	80MB	56.3	Ultra lightweight, prototyping

Image Models

Model	Source	Dimensions	Features
CLIP-ViT-L-14	OpenAI	768	Highest accuracy
CLIP-ViT-B-32	OpenAI	512	Fast
Chinese-CLIP-ViT-H	OFA-Sys	1024	Chinese optimized
SigLIP	Google	1152	Next generation, better performance

---

Cost Calculation

API Model Monthly Cost Estimate

Assumption: 1M documents, average 500 tokens/document

Model	Indexing Cost	Query Cost (100k/month)	Monthly Total
OpenAI small	$10	$2	~$12
OpenAI large	$65	$13	~$78
Cohere v4	$50	$10	~$60
Voyage-3	$30	$6	~$36
Voyage-3-lite	$10	$2	~$12

Open Source Model Deployment Cost

Model	Minimum GPU	Cloud GPU Monthly Cost
BGE-M3	RTX 3090 (24GB)	~$150-300
Qwen3-0.6B	RTX 3060 (12GB)	~$80-150
all-MiniLM	CPU	$0

---

Selection Decision Tree

Start
  │
  ├─ Have GPU?
  │    ├─ No → API models
  │    │         ├─ Tight budget → OpenAI small / Voyage-lite
  │    │         ├─ Best results → Voyage-3-large / Cohere v4
  │    │         └─ General use → OpenAI large
  │    │
  │    └─ Yes → Open source models
  │              ├─ Mainly Chinese → BGE-M3 / Qwen3-Embedding
  │              ├─ Multilingual → BGE-M3
  │              ├─ Ultra lightweight → all-MiniLM / nomic-embed
  │              └─ Images → CLIP / Chinese-CLIP
  │
  └─ Special domain?
       ├─ Code → Voyage-code-3
       ├─ Legal → Voyage-law-2
       └─ Finance → Voyage-finance-2

---

Quick Code Examples

Local Model (Recommended for Beginners)

from sentence_transformers import SentenceTransformer

# Load model (auto-downloads on first use)
model = SentenceTransformer('BAAI/bge-large-zh-v1.5')

# Single encoding
text = "This is a sample text"
embedding = model.encode(text).tolist()

# Batch encoding (recommended)
texts = ["Text 1", "Text 2", "Text 3"]
embeddings = model.encode(texts, batch_size=32, normalize_embeddings=True)

API Model

from openai import OpenAI

client = OpenAI()  # Requires OPENAI_API_KEY

response = client.embeddings.create(
    model="text-embedding-3-small",
    input=["Text 1", "Text 2"]
)
embeddings = [item.embedding for item in response.data]

---

Detailed Integration Guides

See verticals/ directory for detailed integration for each model (registration, installation, code examples):

API Models

• openai.md - OpenAI text-embedding-3
• cohere.md - Cohere Embed v4
• voyage.md - Voyage AI (general/code/legal/finance)
• aliyun.md - Alibaba Cloud DashScope

Open Source Models

• bge.md - BGE series (BAAI)
• qwen-embedding.md - Qwen3-Embedding (Alibaba open source)
• jina.md - Jina Embeddings
• nomic.md - Nomic Embed
• minilm.md - all-MiniLM (lightweight)

Image Models

• clip.md - CLIP / Chinese-CLIP
• siglip.md - SigLIP

---

FAQ

Q: Are more dimensions always better?

Not necessarily. More dimensions mean higher storage costs and slower search. 512-1024 dimensions are usually sufficient; 3072 dimensions are only necessary for extremely high precision scenarios.

Q: Can I mix different models in the same Collection?

No. The same Collection must use the same model, otherwise the vector spaces are different and cannot be compared.

Q: Must queries and documents use the same model?

Yes, queries and documents must be encoded with the same model.

Q: Does OpenAI work well for Chinese?

It works, but BGE-M3 / Qwen3-Embedding perform better for Chinese scenarios.

---

Related Tools

• Batch processing: core:ray
• Chunking strategy: core:chunking
• Index management: core:indexing
• Reranking: core:rerank