Hello, I get this error when using an

At a glance

The community member is experiencing an error when using an index created with the text-embedding-3-large and gpt-4o as the language model. The error message indicates a mismatch in the shapes of the embeddings. The community members suggest the following solutions:

1. Set the Settings.llm and Settings.embed_model in the code to resolve the mismatch.

2. Update the llama-index-llms-openai package and use gpt-4o as the model name without the version.

3. Add the embedding model to the ServiceContext if using that approach.

The community members also discuss the ability to create multiple engines with different configurations, such as lower max_tokens or different temperature, but there is no explicit answer provided.

TTorsten

Hello, I get this error when using an index created with the text-embedding-3-large and gpt-4o as llm:
shapes (1536,) and (3072,) not aligned: 1536 (dim 0) != 3072 (dim 0)
How to fix this?

12 comments

WWhiteFang_Jr

It means that there is some mismatching between embeddings created and newly created embeddings.

I would suggest you add the following at the top:

Plain Text

from llama_index.core import Settings

Settings.llm = your llm instance
Settings.embed_model = your embed model isntance

# then proceed further

TTorsten

But I create the embeddings in a different repository than where my apps run

TTorsten

It actually seems like I cannot use gpt-4o and I was passing model_name and not model before:

Attachment

TTorsten

Which of these models do you recommend and is compatible with text-embedding-3-large

WWhiteFang_Jr

You need to update your llama-index openai package: pip install -U llama-index-llms-openai

WWhiteFang_Jr

Once done just pass gpt-4o as model name only. No need to add model version name

TTorsten

Alright thanks, going to try that

TTorsten

Hmm I still get the shapes error. This is how I set up the model in my app:

storage_context = StorageContext.from_defaults(persist_dir=f"{product_code}_llama")
    index = load_index_from_storage(storage_context)
    llm = OpenAI(temperature=temperature, model=GPT_MODEL, max_tokens=num_outputs)
    service_context = ServiceContext.from_defaults(llm=llm)
    engine = index.as_chat_engine(
        chat_mode="context",
        verbose=True,
        service_context=service_context,
        temperature=temperature,
        system_prompt=prompt,
    )

WWhiteFang_Jr

No need to add service_context. Just define Settings.

If you want to go with service_context. Add the embedding model in there as well

WWhiteFang_Jr

This for settings

TTorsten

Thank you very much, putting everything in the settings worked great!

TTorsten

The only problem I have now is that I want to create multiple engines and for some engines I want a lower max_tokens or a different temperature. How can I tackle this?

Add a reply

Find answers from the community

Hello, I get this error when using an