Is this a bug or my problem? (Embedding model dimension sizes.)

[steveh250]

I fired up LightRAG to use with Ollama and some Granite models (granite-embedding:latest and granite4:latest). Using some code from: https://alain-airom.medium.com/hands-on-experience-with-lightrag-6ecbd3499660 to test it out - the code consistently failed on the embeddings with the following:

ERROR: Embedding func: Error in decorated function for task 139862223312224_3071.870411049: Embedding dimension mismatch detected: total elements (384) cannot be evenly divided by expected dimension (1024).
ERROR: Traceback (most recent call last):
File "/home/ubuntu/MAF/LightRAG/lightrag/lightrag.py", line 1895, in process_document
await asyncio.gather(*first_stage_tasks)
File "/home/ubuntu/MAF/LightRAG/lightrag/kg/nano_vector_db_impl.py", line 124, in upsert
embeddings_list = await asyncio.gather(*embedding_tasks)
File "/home/ubuntu/MAF/LightRAG/lightrag/utils.py", line 503, in call
result = await self.func(*args, **kwargs)
File "/home/ubuntu/MAF/LightRAG/lightrag/utils.py", line 1016, in wait_func
return await future
File "/home/ubuntu/MAF/LightRAG/lightrag/utils.py", line 720, in worker
result = await asyncio.wait_for(
File "/opt/conda/lib/python3.10/asyncio/tasks.py", line 445, in wait_for
return fut.result()
File "/home/ubuntu/MAF/LightRAG/lightrag/utils.py", line 511, in call
raise ValueError(
ValueError: Embedding dimension mismatch detected: total elements (384) cannot be evenly divided by expected dimension (1024).

Set the logging mode to debug and I noticed something odd (snippet below):

embedding_func = EmbeddingFunc(embedding_dim=384, func=functools.partial(EmbeddingFunc(embedding_dim=1024, func=<function ollama_embed at 0x7f34367a11b0>, max_token_size=8192, send_dimensions=False, model_name='bge-m3:latest'), embed_model='granite-embedding:latest', base_url='http://localhost:11434/'), max_token_size=None, send_dimensions=False, model_name=None),
embedding_token_limit = None,
embedding_batch_num = 10,
embedding_func_max_async = 8,
embedding_cache_config = {'enabled': False, 'similarity_threshold': 0.95, 'use_llm_check': False},

The Granite embedding model has an embedding length of 384, after a bit of digging and finding the utils.py where the EmbeddingFunc is I enlisted Claude's help and it suggested the following (I'm not familiar enough with the code or a good enough developer to comment):

O.K. so I thought I would just use the bge model with 1024 embedding dimensions and it worked. In short it seems that only embedding models with 1024 dimensions are supported or there's possibly a bug - am I missing something?

Thanks a bunch.

Cheers - Steve