[Nantha kumar (Nanduu24)]

Motivation and Context

Milvus is one of the most widely used open-source vector databases, but on the Python side it was only available through the deprecated MemoryStore API (connectors/memory_stores/milvus). Every other major backend — Postgres, Qdrant, Redis, Pinecone, Weaviate, Chroma, Faiss, SQL Server, Azure AI Search — has been ported to the modern VectorStore / VectorStoreCollection abstraction, but Milvus had not.

This has been an open, unaddressed request: #13391 ("Why is the Milvus connector not impl VectorStore?"), where another user reported trying to build their own and finding it non-trivial. This PR closes that gap.

Resolves #13391

Description

Adds a new semantic_kernel/connectors/milvus.py with MilvusStore, MilvusCollection and MilvusSettings, following the same structure as the existing qdrant.py / postgres.py connectors:

• Schema from the data model — the key field becomes the primary key (INT64/VARCHAR), data fields become typed scalar fields, and vector fields become FLOAT_VECTOR fields, each with an index. The Milvus dynamic field is enabled so properties outside the definition are preserved on round-trip.
• Mappings — DistanceFunction → Milvus metric types (COSINE/L2/IP), IndexKind → Milvus index types (HNSW/FLAT/IVF_FLAT/DISKANN/AUTOINDEX).
• Full surface — ensure_collection_exists / collection_exists / ensure_collection_deleted, upsert, get, delete, and vector search. Lambda/string filters are translated into Milvus boolean expressions via _lambda_parser (==, !=, <, <=, >, >=, in, not in, and/or/not, chained comparisons).
• Uses the sync MilvusClient wrapped with asyncio.to_thread so it works consistently across the supported pymilvus range (the MilvusClient API requires pymilvus >= 2.4, which the existing shared milvus extra already resolves to — happy to bump the extra floor from 2.3 to 2.4 if you'd prefer it made explicit).
• Registered in connectors/memory.py (lazy import) and connectors/memory.pyi (type stub), mirroring the other connectors.

Testing

• Added tests/unit/connectors/memory/test_milvus.py (28 tests): ingestion, semantic search, filter translation, serialization/deserialization, component lifecycle, and schema creation (hermetic — fake client, no live DB).
• Verified end-to-end against a real Milvus Lite database (create → index → upsert → semantic search → metadata round-trip → filtered search → get → cross-instance persistence → delete), all passing.
• ruff format --check, ruff check, and mypy are clean on the changed files; uv lock --locked is unaffected.

I'm also happy to follow up with a concept sample under samples/concepts/memory/ and a docs entry if that would help.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Allyson (allyson-tcesp)]

Well done! Thx

Is this working with the new Microsoft Agent Framework?

[Nantha kumar (Nanduu24)]

Thanks Allyson (@allyson-tcesp)! This PR targets Semantic Kernel's VectorStore. For the new Microsoft Agent Framework I've just opened the counterpart — a Milvus vector store (MilvusStore/MilvusCollection) — in microsoft/agent-framework#9346, so Milvus will be available in both. Appreciate the interest!

[Nantha kumar (Nanduu24)]

Hi Eduard van Valkenburg (@eavanvalkenburg) — since you own the Python memory connectors, could you review this when you get a moment? It adds a Milvus VectorStore connector (connectors/milvus.py) on the current vector-store base, resolving #13391, with unit tests, mirroring the existing connector pattern. CI is green. (And thanks again for the help on the Milvus work over in agent-framework.)