[Question] : How can I specify embedding dimension with DocumentSummaryIndex?

[keonho-kim]

Question Validation

• I have searched both the documentation and discord for an answer.

Question

Hello, guys.

While dealing with DocumentSummaryIndex, I found that a param, embed_dim in VectoreSotreIndex is not supported.

In this case, how can I control embedding dimension? - I think it would be troublesome when I working with several vectorstores, which have different embedding dimension from DocumentSummaryIndex.

Any ideas?

[dosubot]

Hello, @keonho-kim! I'm here to help you with any bugs, questions, or contributions you may have. Let's tackle this challenge together!

To specify the embedding dimension with DocumentSummaryIndex, you should use the embed_model parameter in the DocumentSummaryIndexEmbeddingRetriever class. The embed_model should be an instance of BaseEmbedding that supports the desired embedding dimension.

Here is an example of how you can specify the embedding dimension:

from llama_index.core.indices.document_summary.base import DocumentSummaryIndex
from llama_index.core.indices.document_summary.retrievers import DocumentSummaryIndexEmbeddingRetriever
from llama_index.core.base.embeddings.base import BaseEmbedding

# Assuming you have a custom embedding model that supports the desired embedding dimension
class CustomEmbeddingModel(BaseEmbedding):
    def __init__(self, embed_dim: int):
        self.embed_dim = embed_dim
        # Initialize other necessary components

    def get_agg_embedding_from_queries(self, queries: List[str]) -> List[float]:
        # Implement the method to return embeddings of the specified dimension
        pass

# Create an instance of your custom embedding model with the desired dimension
custom_embed_model = CustomEmbeddingModel(embed_dim=128)

# Create an instance of DocumentSummaryIndex
document_summary_index = DocumentSummaryIndex(...)

# Create an instance of DocumentSummaryIndexEmbeddingRetriever with the custom embedding model
retriever = DocumentSummaryIndexEmbeddingRetriever(
    index=document_summary_index,
    embed_model=custom_embed_model
)

In this example, CustomEmbeddingModel is a placeholder for your actual embedding model that supports the desired embedding dimension. You need to implement the get_agg_embedding_from_queries method to return embeddings of the specified dimension. Then, you can pass this custom embedding model to the DocumentSummaryIndexEmbeddingRetriever to use it for retrieval ^[1].

To continue talking to Dosu, mention @dosu.

[logan-markewich]

Embed dim goes to the vector store, not the index

Whether your need it or not depends on the vector db you ate using

[keonho-kim]

What I'm confusing is that, what DocumentSumamryIndex returns.

Is it docstore? then what 'embed_summaries' means?

here is what I want to do, but I'm not sure where DocumentSumamryIndex goes. (Now I'm guessing that it is 'docstore')

from llama_index.core.ingestion import IngestionPipeline

pipeline = IngestionPipeline(
          transformations=transformations,
          vector_store=self._vector_store,
          docstore=self._docstore,
          docstore_strategy=docstore_strategy,
          **kwargs,
      )

[logan-markewich]

document summary index is not related to the code you posted?