Summary

Fixes #3970

Upgrades the MongoDB vector store implementation from the deprecated knnBeta/knnVector index syntax to the stable GA vectorSearch format, as documented in the MongoDB Atlas migration guide.

Changes

• VECTOR_TYPE: Changed from "knnVector" (deprecated) to "vector" (GA)
• Index creation (create_col): Replaced legacy mappings definition with type="vectorSearch" and fields list format using numDimensions instead of dimensions
• Search accuracy (search): Increased numCandidates from limit (identical to limit, poor recall) to limit * 20 (recommended 10-20x multiplier for HNSW indexes)

Before (deprecated)

SearchIndexModel(
    name=self.index_name,
    definition={
        "mappings": {
            "dynamic": False,
            "fields": {
                "embedding": {
                    "type": "knnVector",
                    "dimensions": self.embedding_model_dims,
                    ...
                }
            },
        }
    },
)

After (GA)

SearchIndexModel(
    name=self.index_name,
    type="vectorSearch",
    definition={
        "fields": [
            {
                "type": "vector",
                "path": "embedding",
                "numDimensions": self.embedding_model_dims,
                ...
            }
        ]
    },
)

Test plan

• Verify index creation uses type="vectorSearch" with fields list format
• Verify field definition uses numDimensions instead of dimensions
• Verify numCandidates is set to limit * 20 for improved recall
• Verify no knnVector references remain in the codebase