[哈基米 (BlueX888)]

Motivation and Context

FaissCollection._inner_upsert reads the record key by the declared key field name, but records reach it already serialized. The key name in that serialized mapping depends on the serializer: the dict path writes the key's storage_name, while a pydantic model_dump() or a custom to_dict() may write the logical name.

So a definition whose key field declares storage_name fails every upsert of a plain dict record:

VectorStoreOperationException: Error upserting record(s) into collection 'test-alias': 'id'

InMemoryCollection already handles both shapes by using the key name recorded during serialization; FaissCollection overrides _inner_upsert and dropped that logic.

Description

• Read the key name the serialization recorded when it is available (getattr(record, "key_field_name", ...)), otherwise fall back between the storage name and the logical name, mirroring InMemoryCollection._inner_upsert.
• Read the vector under either its storage name or its logical name, so a vector field that declares storage_name works when the record is serialized with logical names, as model_dump() does (second commit).
• Return an empty result when the target index holds no vectors, instead of asking faiss for min(top, 0) neighbours, which fails with a bare assert k > 0. This is the state right after ensure_collection_exists(), and after deleting the last record (third commit).
• Keep filter validation on that empty path: the filter is still resolved before returning, so an invalid filter raises as it does for a populated index.
• Add regression tests with a key field that declares storage_name, parametrized over dict and pydantic records, plus a custom to_dict() that emits logical names, and one test for a pydantic model whose vector field declares storage_name. The parametrized test asserts the returned key, the index key map, inner storage and the search round trip; the to_dict test asserts the returned key and the index key map; the vector test asserts the returned key, the index size, the indexed vector itself and the search round trip. Two more tests cover searching a fresh collection and searching after the last record is deleted; a third asserts that an invalid filter on an empty collection still raises.

Verification

• The new key test fails on main with KeyError: 'id' at faiss.py:183, and passes with this change.
• python/tests/unit/connectors/memory/test_faiss.py (+ test_in_memory.py): 109 passed. On main, the key test fails with KeyError: 'id', the vector-alias test with not enough values to unpack (expected 2, got 1), and both empty-index tests with VectorSearchExecutionException from faiss' assert k > 0.
• ruff check and ruff format --check are clean on both changed files.
• I ran the suite under pydantic 2 and faiss-cpu on Python 3.12 (Windows); the repo's other memory-connector tests are unaffected.

Known limitations, unchanged by this PR: a pydantic model that sets serialize_by_alias=True with a serialization_alias for the vector field that matches neither the logical name nor storage_name still fails, and a definition where one field's logical name equals another field's storage_name resolves to the other field. Both predate this change. Also pre-existing and left alone: FaissCollection._inner_delete keeps the original index offsets after remove_ids compacts positions, so deleting records that are not at the tail can leave the key map misaligned with the index; the delete tests here only remove the last record.

AI disclosure: analysis, implementation and test runs were assisted by Claude Code; I reviewed the change and can explain it.

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 insufficient remaining quota.