Fix OpenAI generic options merging for image, audio, embedding and moderation

[1233567890]

我的代码

	@Autowired
	private OpenAiEmbeddingModel model;
	@Test
	public void ss(){
		EmbeddingResponse floats = model.call(new EmbeddingRequest(List.of("发动机"), EmbeddingOptions.builder().model("bge-multilingual-gemma2").build()));
		System.out.println(floats);
	}

pom配置

<dependency>
			<groupId>org.springframework.ai</groupId>
			<artifactId>spring-ai-starter-model-openai</artifactId>
			<version>2.0.0-M6</version>
		</dependency>

application.yml结构

ai:
    openai:
      base-url: http://192.168.10.233:3000/v1
      api-key: sk-NV44FDGiuspxz27Z1eZqaR2l4YIXI1X1pHgGR5w12i2SI3vS

我确认把配置改成

ai:
   openai:
     embedding:
       model: bge-multilingual-gemma2
     base-url: http://192.168.10.233:3000/v1
     api-key: sk-NV44FDGiuspxz27Z1eZqaR2l4YIXI1X1pHgGR5w12i2SI3vS

这样可以正常工作
如果不修改配置，错误信息如下

com.openai.errors.InternalServerException: 503: No available channel for model text-embedding-ada-002 under group default (distributor) (request id: 202605150707403555553008268d9d6muyPcHkN)

at com.openai.errors.InternalServerException$Builder.build(InternalServerException.kt:95)
at com.openai.core.handlers.ErrorHandler$errorHandler$1.handle(ErrorHandler.kt:79)
at com.openai.core.handlers.ErrorHandler$errorHandler$1.handle(ErrorHandler.kt:40)
at com.openai.services.blocking.EmbeddingServiceImpl$WithRawResponseImpl.create(EmbeddingServiceImpl.kt:80)
at com.openai.services.blocking.EmbeddingServiceImpl.create(EmbeddingServiceImpl.kt:44)
at com.openai.services.blocking.EmbeddingService.create(EmbeddingService.kt:33)
at org.springframework.ai.openai.OpenAiEmbeddingModel.lambda$call$2(OpenAiEmbeddingModel.java:222)
at io.micrometer.observation.Observation.observe(Observation.java:629)
at org.springframework.ai.openai.OpenAiEmbeddingModel.call(OpenAiEmbeddingModel.java:221)
at cn.yuanshivr.chat_example_manager.ChatExampleManagerApplicationTests.ss(ChatExampleManagerApplicationTests.java:90)

[DragonFSKY]

Hi, I looked into this and I think this is a valid issue.

From the current code path, OpenAiEmbeddingModel already passes embeddingRequest.getOptions() into OpenAiEmbeddingOptions.Builder.merge(...). However, merge(...) currently only copies fields when the runtime options are OpenAiEmbeddingOptions. When the request uses the generic EmbeddingOptions.builder(), common fields such as model and dimensions are not merged, so the default OpenAI embedding model is still used.

This seems different from some other embedding model implementations, where common EmbeddingOptions fields are merged before provider-specific fields.

I would be happy to work on a focused fix for this. My plan would be to merge the common EmbeddingOptions#getModel() and getDimensions() fields while keeping OpenAI-specific fields, including deployment name semantics, unchanged.

[sdeleuze]

Please test my fix, be aware there will be a follow-up for fixing OpenAI chat options.