Introduction
text-embedding-3-large represents documents for multilingual retrieval with 3,072 dimensions by default. Its output can be shortened to fit a smaller vector index. Use 3-large when you can evaluate retrieval on the languages and document types in your collection, and when the extra representation size is justified by the quality of the returned matches.
Key features & highlights
The multilingual focus of 3-large is useful when questions and source documents do not use the same language. Evaluate that behavior with examples from your own corpus rather than assuming every translation or specialist term will retrieve equally well. The generated vectors describe semantic relationships; they do not contain a citation, a confidence guarantee or a final answer.
The dimensions option in 3-large can reduce the default representation, including to 1,024 dimensions. A two-passage request at that length and a separate default-length request were verified through the integration. Output length is a storage and retrieval-quality decision, while the measured input token count remains the billing basis for a completed batch.
Request parameters
| Parameter | Required | Description |
|---|---|---|
dimensions | No | Choose 1–3,072 dimensions for 3-large; a 1,024-dimensional request was verified with a two-passage batch. |
embedding_input | Yes | For 3-large, provide the passages themselves rather than instructions to answer a question; each batch result retains its input index. |
Pricing
$0.000026 / 1,000 tokens
Use cases
For a multilingual product knowledge base with 3-large, retrieve source passages for questions written in the customer's language. Keep language, product edition and document revision as separate metadata so an apparently relevant older manual does not displace the current instructions.
For a research collection indexed with 3-large, rank passages from technical documents and supply their original sources to a downstream analysis model. Preserve document boundaries and reference IDs so the final explanation can be checked against the retrieved evidence.
Prompt guide & pro tips
When testing 3-large, build a small set of answerable questions in the languages your users actually use, together with the passages that should be found. Include abbreviations, near-duplicate documents and questions with no suitable source. Compare retrieval results at 3,072 and 1,024 dimensions before choosing the index layout; using the same vector length alone does not make representations from different models interchangeable.
Send 3-large one passage or a homogeneous batch of strings or token arrays. The platform accepts up to 128 items, each within 8,192 tokens, and no more than 300,000 tokens across the request; token arrays are limited to 2,048 entries. Check the returned indexes before storing the vectors, use the same model and dimensions for queries, and rebuild a document's vector after its content changes.
{
"model": "openai/text-embedding-3-large/embedding",
"input": [
"A notebook about coastal birds.",
"An index of harbor observations."
]
}
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "openai/text-embedding-3-large/embedding",
"input": {
"embedding_input": "The customer asks for a refund after being billed twice."
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.


