[ninglinLiu]

BaiduVectorDB.query()/aquery() passed Mochow's raw distance straight into VectorStoreQueryResult.similarities. On the store's default L2 metric that distance is lower-is-closer, so the best match got the lowest similarity and similarity cutoffs / score-based reranking ran backwards. COSINE and IP distances already grow with similarity (Baidu VDB docs).

The store now remembers the table's configured metric and maps L2 distances to 1 / (1 + distance) (same conversion mem0 adopted for its Baidu provider in mem0ai/mem0#6435); COSINE/IP values pass through unchanged. Result order is unchanged.

• Adds mocked-table tests for L2 (closest scores highest) and COSINE/IP pass-through; the L2 test fails on main.
• Bumps the package to 0.8.1.

Fixes #23736

AI disclosure: I used an AI assistant to help locate this and draft the change. I reviewed the diff, checked the metric semantics against Baidu's docs, and ran the package tests plus ruff check/ruff format --check locally.