Spring-ai Possible bug found in TokenCountBatchingStrategy, affecting vector stores

Hello,

While testing the latest build for spring-ai-elasticsearch-store I found myself unable to index documents, getting the following error:

failed to parse: The [dense_vector] field [embedding] in doc [document with id '5e09ac56-e171-4075-a597-e93755fa63a1'] has a different number of dimensions [0] than defined in the mapping [1536]

after some debugging I found out that most of the documents I was trying to index had the embedding field empty, and after some more debugging I have a suspicion that this code might be faulty:

https://github.com/spring-projects/spring-ai/blame/9f33e326b1fa7087c21b2d55d9f147db94332001/spring-ai-core/src/main/java/org/springframework/ai/embedding/TokenCountBatchingStrategy.java#L124

if (currentSize + tokenCount > maxInputTokenCount) {
    batches.add(currentBatch);
    currentBatch.clear();
    currentSize = 0;
}

currentBatch.clear() is also clearing the lists referenced in batches, which gets emptied after every cycle, so only the last few documents actually get to stay in batches, so only those documents are referenced when it's time to fill the embeddings field.

Let me know if more details are needed, happy to help.