[Documentation request] Multi-index for hybrid search

[marquesafonso]

Hi. I really appreciate your project and paper on BMX, as well as the direct integration with HF.

I would like to use it for some projects and found that documentation is lacking on MultiIndex.
I would be happy to help on this if I can get some guidance from your side.

The general idea of what I am trying to achieve is found in the code below:

from baguetter import MultiIndex, BMXSparseIndex, USearchDenseIndex
from baguetter.fuser.config import FuserConfig
import numpy as np

def get_bmx_results(doc_ids:list, docs:list, docs_vectors:np.ndarray, query:str):
    alpha = 0.5
    IndexConfig = [BMXSparseIndex(index_name="bmx"), USearchDenseIndex(index_name="usearch", embedding_dim=256)]
    index = MultiIndex(indices=IndexConfig, fuser_config=FuserConfig(weights=[alpha, 1-alpha], algorithm="weighted"))
    index.add_many(keys=doc_ids, values=list(zip(docs,docs_vectors)))
    res = index.search(query=query)
    return res

I understand that values is of type str | np.ndarray. The docs and docs_vectors (embeddings) are of the same size and the embeddings have been pre-computed, which is why I would not like to pass in an embed_fn, as those computations had already been done. I tried zip to no avail.

Would you be able to help me on this?

Thank you for your time and sharing this project openly.

[juliuslipp]

Hey @marquesafonso,

I think easiest way would be to add the values separately to each index beforehand. e.g.

sparse = BMX(...
dense = Usearch(...

sparse.add_many(..., values=texts)
dense.add_many(..., values=vectors)

What add_many on multi-index does, is calling add_many of each index with the same values you passed to multi.add_many.

Does that work for you?

[marquesafonso]

Hi @juliuslipp,

Thank you for the reply and apologies for taking a while to respond.

I am trying to evaluate the usage of a hybrid retriever (sparse + dense) and would like to offload index computation to hugging face datasets (such that these computations are only done once).

I am having trouble saving the index in a windows machine. I get the following error:

OSError: [WinError 123] The filename, directory name, or volume label syntax is incorrect: 'C:\\Users\\user\\.cache\\baguetter\\repository/C:'

I believe this arises from:

def save(self, path: str) -> str:
        repository = LocalFileRepository()
        directory = path.rsplit("/", 1)
        if len(directory) > 1:
            repository.mkdirs(directory[0], exist_ok=True)
        path = self._save(path=path, repository=repository)
        return repository.info(path)["name"]

In particular the .rsplit used with "/" should cause problems. Maybe it could be more robust to use the os.path module (and methods such as dirname or basename)?

Still on this subject, perhaps the .push_to_hub() method could work without the path argument? This because when we call it we are already calling it on an index object (in-memory).

Could you help me out on this?

[juliuslipp]

Hey @marquesafonso, unfortunately we won't be able to support windows. If there's any PR that helps with fixing this issue, I'm happy to take a look though!

[marquesafonso]

Hi @juliuslipp, quick message just to directly point that I opened a PR based on our conversation above. If you could take a look when you have some time it would be greatly appreciated (by myself and possibly other windows users)!