AI functions

Embedding functions

openai_embedding

• Added in v2.5.0.
• Since v2.6.0, openai_embedding function has been updated to accept a single config argument as a constant JSONB value containing the embedding configuration, instead of separate api_key and model arguments.

Get the text embedding vector for a given text. The config argument must be a constant JSONB value containing the embedding configuration.

Syntax

openai_embedding ( config constant jsonb, input text ) → real[]

The config parameter is a JSONB object that supports the following options:

Option	Required	Description
model	Yes	The embedding model to use (e.g., text-embedding-3-small)
api_key	No	OpenAI API key
org_id	No	OpenAI organization ID
project_id	No	OpenAI project ID
api_base	No	Custom API base URL. Default: https://api.openai.com/v1

Example

openai_embedding('{"model": "text-embedding-3-small", "api_key": "sk-proj-1234567890"}'::jsonb, 'Hello, world!') -> {0.0001, 0.0002, 0.0003, ...}

Retry behavior

If an OpenAI embedding request fails, RisingWave retries it with exponential backoff. The delay starts at 100 milliseconds and is capped at 10 seconds. There is no built-in retry limit. If OpenAI continues returning errors, RisingWave continues retrying until it receives a valid embedding response.

Metrics

RisingWave exposes the following Prometheus metrics for openai_embedding. Each metric has model and api_base labels.

Metric	Description
openai_embedding_success_count	Total number of successful embedding requests.
openai_embedding_failure_count	Total number of failed request attempts. Retries are counted as separate failed attempts.
openai_embedding_latency	Latency in seconds for successful request attempts.
openai_embedding_input_rows	Total number of non-null, non-empty input rows submitted for embedding.
openai_embedding_input_row_word_count	Distribution of word counts for non-null, non-empty input rows.
openai_embedding_inflight_requests	Number of embedding requests currently in flight.
openai_embedding_inflight_rows	Number of input rows in in-flight embedding requests.

For information about accessing Prometheus metrics, see Prometheus metrics reference.