[[["易于理解","easyToUnderstand","thumb-up"],["解决了我的问题","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["很难理解","hardToUnderstand","thumb-down"],["信息或示例代码不正确","incorrectInformationOrSampleCode","thumb-down"],["没有我需要的信息/示例","missingTheInformationSamplesINeed","thumb-down"],["翻译问题","translationIssue","thumb-down"],["其他","otherDown","thumb-down"]],["最后更新时间 (UTC):2025-09-05。"],[],[],null,["# Vector indexing best practices\n\n| **PostgreSQL interface note:** The examples in this topic are intended for GoogleSQL-dialect databases. This feature doesn't support PostgreSQL interface.\n\n\u003cbr /\u003e\n\n\n| **Note:** This feature is available with the Spanner Enterprise edition and Enterprise Plus edition. For more information, see the [Spanner editions overview](/spanner/docs/editions-overview).\n\n\u003cbr /\u003e\n\nThis page describes vector indexing best practices that optimize your\n[vector indexes](/spanner/docs/vector-indexes) and improve\n[approximate nearest neighbor (ANN) query results](/spanner/docs/find-approximate-nearest-neighbors#query-vector-embeddings).\n\nTune the vector search options\n------------------------------\n\nThe most optimal values for your vector index options depend on your use case,\nvector dataset, and on the query vectors. You can set and tune these values\nby creating a new vector index and setting the [`index_option_list`](/spanner/docs/reference/standard-sql/data-definition-language#index_option_list)\nin the `CREATE VECTOR INDEX` statement. You might need to perform iterative\ntuning to find the best values for your specific workload.\n\nHere are some helpful guidelines to follow when picking appropriate values:\n\n- `tree_depth` (tree level): If the table you're indexing has fewer than 10\n million rows, use a `tree_depth` of `2`. Otherwise, a `tree_depth` of `3`\n supports tables of up to about 10 billion rows.\n\n- `num_leaves`: Use the square root of the number of rows in the dataset. A\n larger value can increase vector index build time. Avoid setting `num_leaves`\n larger than the `table_row_count` divided by 1000 as this results in overly\n small leaves and poor performance.\n\n- `num_leaves_to_search`: This option specifies how many leaf nodes of the index\n are searched. Increasing `num_leaves_to_search` improves recall but also\n increases latency and cost. We recommend using a number that is 1% the total\n number of leaves defined in the `CREATE VECTOR INDEX` statement as the value\n for `num_leaves_to_search`. If you're using a filter clause, increase\n this value to widen the search.\n\nIf acceptable recall is achieved, but the cost of querying is too high,\nresulting in low maximum QPS, try increasing `num_leaves` by following these\nsteps:\n\n1. Set `num_leaves` to some multiple k of its original value (for example, `2 * sqrt(table_row_count)`).\n2. Set `num_leaves_to_search` to be the same multiple k of its original value.\n3. Experiment with reducing `num_leaves_to_search` to improve cost and QPS while maintaining recall.\n\nImprove recall\n--------------\n\nTo improve recall, consider tuning the `num_leaves_to_search` value or\nrebuilding your vector index.\n\n### Increase the `num_leaves_to_search` value\n\nIf the `num_leaves_to_search` value is too small, you might find it more\nchallenging to find the nearest neighbors for some query vectors. Creating a new\nvector index with an increased `num_leaves_to_search` value can help improve\nrecall by searching more leaves. Recent queries might contain more of these\nchallenging vectors.\n\n### Rebuild the vector index\n\nThe tree structure of the vector index is optimized for the dataset at the time\nof creation, and is static thereafter. Therefore, if significantly different\nvectors are added after creating the initial vector index, then the tree\nstructure might be sub-optimal, leading to poorer recall.\n\nTo rebuild your vector index without downtime:\n\n1. Create a new vector index on the same embedding column as the current vector index, updating parameters (for example, `OPTIONS`) as appropriate.\n2. After the index creation completes, use the [`FORCE_INDEX` hint](/spanner/docs/secondary-indexes#index-directive) to point at the new index to update the vector search query. This ensures that the query uses the new vector index. You might also need to retune `num_leaves_to_search` in your new query.\n3. Drop the outdated vector index.\n\nWhat's next\n-----------\n\n- Learn more about Spanner [vector indexes](/spanner/docs/vector-indexes).\n\n- Learn more about Spanner [approximate nearest neighbors](/spanner/docs/find-approximate-nearest-neighbors).\n\n- Learn more about the [GoogleSQL `APPROXIMATE_COSINE_DISTANCE()`, `APPROXIMATE_EUCLIDEAN_DISTANCE()`, `APPROXIMATE_DOT_PRODUCT()`](/spanner/docs/reference/standard-sql/mathematical_functions) functions.\n\n- Learn more about the [GoogleSQL `VECTOR INDEX` statements](/spanner/docs/reference/standard-sql/data-definition-language#vector_index_statements)."]]