AttributeError: 'HuggingFaceEmbedding' object has no attribute '_model'

[JoepdeJong]

Due to recent updates in llama_index, the code breaks when using multiple workers (run-llama/llama_index#13497).

This can be temporary fixed by setting:

llama-index = "==0.10.30"
...
llama-index-core = "==0.10.30"

In the Pipfile.

[jonfairbanks]

Thanks for opening this. I'll also keep an eye on the bug thread as this was definitely working with values > 1 before the update.

[JoepdeJong]

I did some debugging. It seems that _model is missing the second time the VectorStoreIndex is used.

This index is initialized here:

local-rag/utils/llama_index.py

Line 172 in 72f2550

index = create_index(_documents)

If you debug after this line, or print e.g. the following

print(hasattr(index._embed_model, "_model"))

you will see that the model is missing the second time: on file upload it exists, on chat, it isnt there.

It seems to be a problem in the caching:

local-rag/utils/llama_index.py

Line 114 in 72f2550

@st.cache_data(show_spinner=False)

After removing @st.cache_data(show_spinner=False), it seems to work again.

Edit:
It is clearly the caching, even running

        index = create_index(_documents)
        print(
            hasattr(index._embed_model, "_model"),
        )
        index2 = create_index(_documents)
        print(
            hasattr(index2._embed_model, "_model"),
        )

loses the model attribute.