Liu Liu created FLINK-41001:
-------------------------------
Summary: OpenAI model provider rejects compatible string input
types
Key: FLINK-41001
URL: https://issues.apache.org/jira/browse/FLINK-41001
Project: Flink
Issue Type: Bug
Components: Library / Machine Learning
Affects Versions: 2.3.0, 2.2.0, 2.1.0, 2.4.0
Reporter: Liu Liu
The OpenAI model provider validates its model input schema against an exact
nullable, unbounded {{STRING}} type. Consequently, models declaring
{{{}CHAR(n){}}}, bounded {{{}VARCHAR(n){}}}, or {{STRING NOT NULL}} inputs fail
during prediction planning, although the provider supports their string
representation.
For example,
{code:java}
CREATE MODEL chat_model
INPUT (prompt CHAR(5) NOT NULL)
OUTPUT (answer STRING)
WITH (
'provider' = 'openai',
'endpoint' = 'https://example.test/chat/completions',
'api-key' = 'dummy',
'model' = 'example'
);
CREATE TEMPORARY VIEW prompts AS
SELECT 'hello' AS prompt;
EXPLAIN PLAN FOR
SELECT * FROM TABLE(
ML_PREDICT(TABLE prompts, MODEL chat_model, DESCRIPTOR(prompt))
); {code}
Planning fails with
{code:java}
input column prompt should be STRING, but is a CHAR(5) NOT NULL. {code}
The shared validation affects OpenAI models through SQL and Table API.
Input validation should accept CHAR and VARCHAR regardless of their length or
nullability, while retaining the existing output schema, column count, and
metadata constraints.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
