Amanai Docs

Embeddings

OpenAI-compatible text embeddings for search, clustering, recommendations, and retrieval.

Overview

Use POST /v1/embeddings with the same Bearer API key and OpenAI-compatible SDK contract used by your other Amanai integrations. The endpoint accepts one string or a batch of up to 64 strings and returns one vector for each input item.

Discovery is typed. List the models available to your key with GET /v1/models?type=embedding. Do not select chat models on this endpoint.

Embedding models

Model	Dimensions	Languages	Best fit
BGE Small EN `amanai/bge-small-en`	384	English	Low-latency indexing
BGE Base EN `amanai/bge-base-en`	768	English	Balanced retrieval
BGE Large EN `amanai/bge-large-en`	1024	English	Higher-quality English retrieval
BGE-M3 `amanai/bge-m3`	1024	Multilingual	Multilingual retrieval
EmbeddingGemma 300M `amanai/embeddinggemma-300m`	768	Multilingual	Multilingual retrieval
Qwen3 Embedding 0.6B `amanai/qwen3-embedding-0.6b`	1024	Multilingual	Multilingual retrieval
voyage-4-lite `amanai/voyage-4-lite`	1024	Multilingual	General text retrieval
voyage-4 `amanai/voyage-4`	1024	Multilingual	General text retrieval
voyage-4-large `amanai/voyage-4-large`	1024	Multilingual	General text retrieval
voyage-code-4 `amanai/voyage-code-4`	1024	Multilingual	Code retrieval
voyage-finance-2 `amanai/voyage-finance-2`	1024	Multilingual	Finance retrieval
voyage-law-2 `amanai/voyage-law-2`	1024	Multilingual	Legal retrieval

Request

bash

curl https://api.amanai.dev/v1/embeddings \
  -H "Authorization: Bearer $AMANAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amanai/bge-m3",
    "input": ["hello world", "semantic search"],
    "encoding_format": "float"
  }'

Field	Type	Behavior
`model`	string	Required embedding model ID from typed discovery.
`input`	string or string[]	One text item or up to 64 text items.
`encoding_format`	string	`float` for numeric vector arrays or `base64` for encoded vectors.
`dimensions`	integer, optional	Voyage 4-series models support 256, 512, 1024, or 2048. Voyage finance/law support 1024 only. Other embedding models keep their advertised native dimension.
`input_type`	string, optional	Voyage accepts `query` or `document`. Use the matching type for search queries and indexed documents; omit this field for legacy embedding models.
`user`	string, optional	A bounded end-user identifier for the request.

Voyage request example

bash

curl https://api.amanai.dev/v1/embeddings \
  -H "Authorization: Bearer $AMANAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amanai/voyage-4",
    "input": ["A document about semantic search"],
    "input_type": "document",
    "dimensions": 1024
  }'

Response

bash

{
  "object": "list",
  "model": "amanai/bge-m3",
  "data": [
    {"object": "embedding", "index": 0, "embedding": [0.012, -0.034, "..."]},
    {"object": "embedding", "index": 1, "embedding": [0.021, -0.018, "..."]}
  ],
  "usage": {"item_count": 2, "dimensions": 1024, "encoding_format": "float"}
}

Keep the model ID and dimensions with your stored vectors. Queries and indexed documents must use the same model and vector dimension.

Batching and retrieval

✓Batch related inputs to reduce request overhead, while respecting the endpoint's accepted payload size.

✓Normalize your application text consistently before indexing and querying.

✓Use amanai/bge-m3 or another multilingual model when Indonesian, Malay, and English can coexist.

✓Use Rerank after vector retrieval when you need a more precise final ordering.

Billing

BGE, Qwen3, and Voyage embeddings bill tokens with a 250-credit minimum. EmbeddingGemma remains at 1 credit per UTF-8 unit with a 1,000-credit minimum. Local byte counts are not exact token counts. Failed requests before delivery are not charged. See Billing & pricing for current per-model rates.