Embeddings

Turn text or supported vision inputs into vectors for RAG, semantic search, clustering, and recommendations. Start text workloads with text-embedding-v4; use /v1/models and the model detail page for vision embedding models.

Supported models

Model	Dim	Ctx	Price / 1M	Best for
text-embedding-v4	1024	8K	$0.13	General, English-leaning
tongyi-embedding-vision-plus	Model-specific	Model-specific	See model details	Vision embedding and image-text retrieval
tongyi-embedding-vision-flash	Model-specific	Model-specific	See model details	Lower-latency vision embedding
qwen3-vl-embedding	Model-specific	Model-specific	See model details	Multimodal embedding tasks

Example

Python

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://test.sealink.io/v1",
    api_key=os.environ["SEALINK_API_KEY"],
)

# Single string
res = client.embeddings.create(
    model="text-embedding-v4",
    input="SeaLink helps SEA developers ship AI faster."
)
vec = res.data[0].embedding   # 1024-dimensional vector

# Batch (recommended for performance)
texts = ["Doc 1 content", "Doc 2 content", "Doc 3 content"]
res = client.embeddings.create(model="text-embedding-v4", input=texts)
vectors = [d.embedding for d in res.data]

Parameters

Field	Type	Description
model	string	text-embedding-v4, tongyi-embedding-vision-plus, tongyi-embedding-vision-flash, qwen3-vl-embedding
input	string \| string[]	Text to embed; arrays must not be empty
encoding_format	string	float, base64
dimensions	integer	Optional output dimension override when supported; must be greater than 0

Which one?

text-embedding-v4: General-purpose text embedding model for RAG, semantic search, and recommendation baselines. It returns 1024-dimensional vectors.
tongyi-embedding-vision-plus / flash: Use for image, image-text retrieval, and visual similarity tasks. Choose input fields from the model detail page.
qwen3-vl-embedding: Use for multimodal embedding probes; first probes should use one short input.

Performance tips

Prefer batch input (array form) to reduce request overhead and improve throughput.
If you use cosine similarity, L2-normalize before storing so retrieval can use dot product.
Chunk size: 500-800 tokens with 50-100 token overlap is a safe default.
Before launch, evaluate retrieval quality, reranking, and cost with a representative query set.