feat: to support similarity with relevance scores

[cheriL]

Description: to support _similarity_search_with_relevance_scores func in base class VectorStore with milvus.

[cheriL]

fix conflicts and lints.

[cheriL]

From the CI error report, I just realized that the latest code supports multiple searches(I used to use the deprecated one...). So the problem may be more complicated than I thought, and the methods in the base class may not be available.

[zc277584121]

@cheriL thank you for your contribution. can you tell me what 's the issue without this contribution.
As i know ,there seems raise an issue in this code with current langchain_milvus code version:

retriever = vdb.as_retriever(
    search_type='similarity_score_threshold',
    search_kwargs={
        'score_threshold': 0.3,
        'k': 3,
    },
)
res = retriever.invoke("hello world")
print(res)

...
  File "/Users/zilliz/miniforge3/envs/py310/lib/python3.10/site-packages/langchain_core/vectorstores/base.py", line 560, in similarity_search_with_relevance_scores
    docs_and_similarities = self._similarity_search_with_relevance_scores(
  File "/Users/zilliz/miniforge3/envs/py310/lib/python3.10/site-packages/langchain_core/vectorstores/base.py", line 507, in _similarity_search_with_relevance_scores
    relevance_score_fn = self._select_relevance_score_fn()
  File "/Users/zilliz/miniforge3/envs/py310/lib/python3.10/site-packages/langchain_core/vectorstores/base.py", line 448, in _select_relevance_score_fn
    raise NotImplementedError

So I think maybe your contribution want to solve this sort of issue.
can you tell the origin purpose of this contribution. Just want to collect more cases and avoid all the potential issues

[cheriL]

@zc277584121 My problem is similar to your example.
I'm working on building a RAG application using milvus as my vector store. when I expect to get docs and normalized scores by calling _similarity_search_with_relevance_scores, NotImplemented error raised. and I handle it by metric_type outside the function currently.
may also be something wrong with my usage?

[zc277584121]

@cheriL Currently the _similarity_search_with_relevance_scores() and _select_relevance_score_fn() are missing in Milvus class , which causes our issues, I will add them within about one day.
I'm now drafting the code by refering to this PR, but I found there are some potential problem by directly use _euclidean_relevance_score_fn when 'L2' and use _max_inner_product_relevance_score_fn when IP because the formulas in both methods can not gaurantee the so called relevance score are between [0, 1] . So I decided to write are custom formulas for milvus.
On the other hand, I will only consider the dense single vector situation, since other cases are complicated, and leaving them for further consideration.
Feel free to let me know if you have some problem with this design. :)

[cheriL]

I agree that the multi-vector situation is quite complex. Thanks for your work on this project!

[zc277584121]

done: #18
@cheriL FYI

[zc277584121]

so i close this one