[Feature Request] Add runtime filter support to MilvusEmbeddingRetriever

[kukjun]

Problem

MilvusEmbeddingRetriever doesn't support runtime filtering, unlike other Haystack retrievers (Elasticsearch, InMemory, etc.).

Current signature:

def run(self, query_embedding: List[float], top_k: int)

Expected signature:

def run(self, query_embedding: List[float], top_k: int, filters: Optional[Dict[str, Any]] = None)
Example Usage
retriever = MilvusEmbeddingRetriever(document_store=document_store)

filters = {
    "operator": "AND",
    "conditions": [
        {"field": "meta.category", "operator": "==", "value": "diabetes"},
        {"field": "meta.year", "operator": ">=", "value": 2020}
    ]
}


results = retriever.run(query_embedding=embedding, filters=filters)

Why This Matters

• Consistency: All other vector stores support this - Milvus shouldn't be an exception
• Performance: Database-level filtering is far more efficient than post-retrieval filtering in Python
• Interoperability: Enables seamless switching between vector databases
• Use Cases: Multi-tenancy, category filtering, time-based search, etc.

Implementation Approaches

I'm considering two approaches and would appreciate maintainer input:

1. Milvus-Native

Convert Haystack filters to Milvus boolean expressions:

# {"field": "meta.category", "operator": "==", "value": "diabetes"}
# → "category == 'diabetes'"

2. Unified Integration

Align with a standardized filter format that works across multiple vector databases (Milvus, Pinecone, Weaviate, Qdrant), making the entire Haystack ecosystem more consistent.

Technical Details

Milvus supports filtering via Collection.search(expr=...)
pymilvus already provides this functionality
Just need to bridge Haystack's filter format to Milvus expressions

[zc277584121]

add at #70