Initial API POC

[piercefreeman]

Very rough initial sketch of the ORM syntax. This PR is aiming to:

• Establish initial model specification
• Establish initial query builder
• Handle boolean and vector similarity filters

class MyObject(MilvusBase, milvus_client=Milvus()):
    __embedding_dim__ = 128
    __collection_dim__ = "my_collection"

    text: str
    embedding: Embedding
    id: int

    def __init__(self, text: str, embedding: Embedding):
        super().__init__()
        self.text = text
        self.embedding = embedding
        self.id = None

Insertion instructions:

obj1 = MyObject('foo', Embedding([1.0] * 128))
obj1.insert(milvus_client)

Search instructions:

results = session.query(MyObject).filter(MyObject.text == 'bar').order_by_similarity(MyObject.embedding, Embedding([1.0]*128)).limit(2).all()
assert len(results) == 1

[hweller1]

good stuff! tests make usage pretty clear, and i like some of the extensions you've added to the Milvus ORM.

think it would be nice to add a readme to articulate a bit what this does that the pymilvus ORM (or other future vector db ORMs for that matter) do not to make usage a bit more obvious. I also think a Jupyter notebook with some examples would make this thing ✨

[hweller1]

recommend using the setter decorator for these: https://www.geeksforgeeks.org/getter-and-setter-in-python/

[piercefreeman]

@hweller1 These are intended to be used as chaining functions query().filter(XYZ).offset(2) - are you envisioning a use case where people are directly using getters and setters as part of this chain?

[hweller1]

oh i think the usage shouldn't have to change, was just saying that there is a standard way of defining attribute setting. more syntactic than functional

[hweller1]

would augment this with some quick benchmarks. Very simple search task where IVF_FLAT is a better choice than say HNSW or ANNOY

[hweller1]

it could be good to expose the __repr__ of the Collection or CollectionSchema somewhere- the user can also pull it by calling str(collection) or str(collection._schema)

Here and here are the definitions for them, they seems useful

[hweller1]

took me a bit of digging but it's probably worth highlighting that this is the one type conversion you have that Milvus does not support out of the box, and is a major improvement on it's ORM! #numpy is all you need

[hweller1]

when does Milvus return this vs an ordinary QueryResult ?