feat: add multi vector support #11
Conversation
|
Hi @omriel1 thanks for the PR! FYI, the reviewer @zc277584121 is out for a week. Is this PR urgent? Would prefer to wait till the reviewer is back, or we could ask @efriis to take a look first. |
|
Hi! Thanks for the reply. This is important for us (it was delayed due to the repo separation), but I completely understand and agree that we should wait for @zc277584121. But in the meantime, if you or @efriis have any comments, I'd be happy to start addressing them! Thanks! |
| dense_embeddings_func_2, | ||
| ], | ||
| texts=fake_texts, | ||
| connection_args={"uri": "./milvus_demo.db"}, |
There was a problem hiding this comment.
Can this uri be passed in the same way as ut above using temp_milvus_db? Just for the code consistency
| embedding=[FakeEmbeddings(), FakeEmbeddings()], | ||
| index_params=[index_param_1, index_param_2], | ||
| vector_field=["vec_field_1", "vec_field_2"], | ||
| connection_args={"uri": "./milvus_demo.db"}, |
| embedding=[embedding_1, embedding_2], | ||
| texts=fake_texts, | ||
| index_params=[index_param_1, index_param_2], | ||
| connection_args={"uri": "./milvus_demo.db"}, |
| for i, embeddings_func in enumerate(embeddings_functions): | ||
| if not self._get_index(vector_fields[i]): | ||
| try: | ||
| # If no index params, use a default HNSW based one |
There was a problem hiding this comment.
maybe we should keep the comment If no index params, use a default *AutoIndex* based one unchanged
| using=self.alias, | ||
| ) | ||
|
|
||
| # If default did not work, most likely on Zilliz Cloud |
|
|
||
| # If default did not work, most likely on Zilliz Cloud | ||
| except MilvusException: | ||
| # Use AUTOINDEX based index |
|
@omriel1 sorry for delay. thanks for your great contributing. and i have left some small comments here. all others LGTM |
Description
This PR introduces multi-vector support!
@efriis @zc277584121 I've moved and modified the PR previously submitted to the main langhcain repo: langchain-ai/langchain#26500
Milvus 2.4 introduced the option for multi-vector support, which is becoming increasingly popular, especially for use cases like hybrid search (dense + sparse embeddings).
Lately, @ohadeytan introduced the option to use sparse embeddings in this PR.
Additionally, @zc277584121 already introduced the
MilvusCollectionHybridSearchRetriever, which enables hybrid search against pre-defined collections directly via pymilvus.However, this method doesn't take full advantage of the many useful features offered by
langchain_milvuswhen building a collection: automatic schema creation, indexing and search parameter creation etc.This PR intend to make developers life easier, by allowing them to use single-vector or multi-vector with a single
langchaininterface, that create, connect, and searchMilvus. For example, at IBM and IBM Research this feature is requested by many of our developers and researchers and will be very useful for us.Changes
This PR addresses the limitations described above by introducing the following changes:
We are eager to have this merged into
langchain-milvusas we utilize many of langchain features, particularlylangchain-milvus. We want to continue benefiting from the many valuable features this package provides. We'll make any required changes and will be glad to get any guidance to make it happen!Twitter handle: @EliyahuOmri, @ohadeytan