The community members are discussing the context window size of the bge embedding models, which is 512, compared to the default 1024 chunk size in the LlamaIndex. They presume that the LlamaIndex is truncating the input to the first 512 tokens when using these models. The community members suggest that the chunk size should be equal to the max_position_embeddings of the embedding model, and one member suggests adding a warning to notify users of the mismatch between the chunk size and the embedding model.
So the bge embedding models are the leading OS embedding models; however, they have a 512 context window. I can see that the defaults for llama index is 1024. I presume that they are just taking the first 512 tokens when embedding a 1024 chunks size with such a model.
Shouldn't upper bounds of the chunk size be equal to the max_position_embeddings of the embedding model.
ya that's a good suggestion. Right now the default setting is not dynamic based on the model selected. And the setting is mostly optimized for OpenAI embedding / models
