问题：AzureOpenAi模型在达到令牌长度限制时应使用FinishReason.LENGTH。如果我们发送了太多的令牌，那么Azure Open AI会用一个http状态400和一个包含错误代码的json响应体来回答。这个400状态会引发OpenAiClient的getChatCompletions方法抛出一个异常。

这个问题在这里处理：
langchain4j/langchain4j-azure-open-ai/src/main/java/dev/langchain4j/model/azure/AzureOpenAiChatModel.java
第313行到第336行

} catch (HttpResponseExceptionhttpResponseException) {
    logger.info("Error generating response, {}", httpResponseException.getValue());
    FinishReasonexceptionFinishReason = contentFilterManagement(httpResponseException, "content_filter");
    Response<AiMessage> response = Response.from(
        aiMessage(httpResponseException.getMessage()),
        null,
        exceptionFinishReason
    );
    ChatModelErrorContexterrorContext = newChatModelErrorContext(
        httpResponseException,
        modelListenerRequest,
        null,
        attributes
    );

    listeners.forEach(listener -> {
        try {
            listener.onError(errorContext);
        } catch (Exceptione2) {
            logger.warn("Exception while calling model listener", e2);
        }
    });
    returnresponse;
}

辅助类只处理ContentFilter的情况。
langchain4j/langchain4j-azure-open-ai/src/main/java/dev/langchain4j/model/azure/InternalAzureOpenAiHelper.java
第337行到第359行

publicstaticFinishReasoncontentFilterManagement(HttpResponseExceptionhttpResponseException, StringcontentFilterCode) {
    FinishReasonexceptionFinishReason = FinishReason.OTHER;
    if (httpResponseException.getValue() instanceof Map) {
        try {
            Map<String, Object> error = (Map<String, Object>) httpResponseException.getValue();
            ObjecterrorMap = error.get("error");
            if (errorMapinstanceof Map) {
                Map<String, Object> errorDetails = (Map<String, Object>) errorMap;
                ObjecterrorCode = errorDetails.get("code");
                if (errorCodeinstanceof String) {
                    Stringcode = (String) errorCode;
                    if (contentFilterCode.equals(code)) {
                        // The content was filtered by Azure OpenAI's content filter (for violence, self harm, or hate).
                        exceptionFinishReason = FinishReason.CONTENT_FILTER;
                    }
                }
            }
        } catch (ClassCastExceptionclassCastException) {
            logger.error("Error parsing error response from Azure OpenAI", classCastException);
        }
    }
    returnexceptionFinishReason;
}

我认为它还应该管理令牌限制达到的情况，然后这个方法应该被重命名为例如TokenLimitReached。在当前实现中，我们得到了一个TokenLimitReached,但我们失去了达到了限制的事实(即使限制应该由业务应用程序正确管理；-)