向量检索支持SQL全集

更新时间：2026-09-11

本文汇总 Doris 向量检索当前支持和不支持的 SQL 写法，包含建表 DDL、查询语法、多路召回与重排序示例，供开发和测试对照使用。

概览

Doris 是一款集向量检索、全文检索、OLAP 分析功能于一体的数据库，其中存算分离版具有高可用、高性能、低成本的特点。向量检索既可以用标准 SQL 完成，也可以通过 SDK 获得更高 QPS。

功能	说明
向量维度	2-32,768（更高维度也支持）
度量类型	余弦、IP（点积）、L2（欧几里得）、BM25（全文检索）
索引类型	HNSW、DiskANN、Puck（IVF_PQ）
搜索类型	向量检索、标量过滤、全文检索、OLAP 分析
全文检索	支持，BM25 算法
支持的 SDK	标准 SQL、JDBC/ODBC、Python/Go SDK

DDL

索引属性说明

属性	说明
索引名称	如 `idx_ann`，在 `INDEX` 关键字后指定
`algorithm`	向量索引算法，示例中使用 `hnsw`
`dimension`	指定向量维度
`distance`	指定向量距离计算方式，示例中使用 `l2`

Duplicate 表

SQL

1CREATE TABLE `ann_table` (
2  `F_id` int NOT NULL COMMENT 'id',
3  `F_meta_id` char(6) NOT NULL COMMENT 'meta id',
4  `F_content` text NULL COMMENT 'content1',
5  `embedding` array<float> NOT NULL COMMENT 'embedding',
6  `len` INT DEFAULT '1' COMMENT 'meta length',
7  INDEX idx_ann (`embedding`) USING ANN PROPERTIES("algorithm"="hnsw", 'dimension'='4', 'distance'='l2') COMMENT 'hnsw index'
8) ENGINE=OLAP
9DUPLICATE KEY(`F_id`)
10COMMENT 'OLAP'
11DISTRIBUTED BY HASH(`F_id`) BUCKETS 2
12PROPERTIES (
13);

Unique 表

与 Duplicate 表的差别只在 KEY 类型，索引定义方式相同。

SQL

1CREATE TABLE `ann_table_unique` (
2  `F_id` int NOT NULL COMMENT 'id',
3  `F_meta_id` char(6) NOT NULL COMMENT 'meta id',
4  `F_content` text NULL COMMENT 'content1',
5  `embedding` array<float> NOT NULL COMMENT 'embedding',
6  `len` INT DEFAULT '1' COMMENT 'meta length',
7  INDEX idx_ann (`embedding`) USING ANN PROPERTIES("algorithm"="hnsw", 'dimension'='4', 'distance'='l2') COMMENT 'hnsw index'
8) ENGINE=OLAP
9UNIQUE KEY(`F_id`)
10COMMENT 'OLAP'
11DISTRIBUTED BY HASH(`F_id`) BUCKETS 2
12PROPERTIES (
13);

多列索引

支持为多个字段分别创建向量 ANN 索引、文本倒排索引，查询时可组合多个索引实现混合检索，提升多条件查询性能。

SQL

1CREATE TABLE `hybrid_dup_table` (
2  `id` int NOT NULL COMMENT 'id',
3  `name` varchar(50) NOT NULL COMMENT 'name',
4  `age` int NULL COMMENT 'age',
5  `embedding1` array<float> NOT NULL COMMENT 'embedding1',
6  `embedding2` array<float> NOT NULL COMMENT 'embedding2',
7  `embedding3` array<float> NOT NULL COMMENT 'embedding2',
8  `content1` String NOT NULL COMMENT 'content1',
9  `content2` String NOT NULL COMMENT 'content2',
10  INDEX idx_ann1 (`embedding1`) USING ANN PROPERTIES("algorithm"="hnsw", 'hnsw_m'='1000', 'hnsw_efConstruction'='2000', 'dimension'='4', 'distance'='l2') COMMENT 'hnsw index',
11  INDEX idx_ann2 (`embedding2`) USING ANN PROPERTIES("algorithm"="hnsw", 'hnsw_m'='1000', 'hnsw_efConstruction'='2000', 'dimension'='4', 'distance'='l2') COMMENT 'hnsw index',
12  INDEX idx_ann3 (`embedding3`) USING ANN PROPERTIES("algorithm"="hnsw", 'hnsw_m'='1000', 'hnsw_efConstruction'='2000', 'dimension'='4', 'distance'='l2') COMMENT 'hnsw index',
13  INDEX idx_content1 (`content1`) USING INVERTED PROPERTIES("parser" = "english") COMMENT 'inverted index',
14  INDEX idx_content2 (`content2`) USING INVERTED PROPERTIES("parser" = "english") COMMENT 'inverted index'
15) ENGINE=OLAP
16DUPLICATE KEY(`id`)
17COMMENT 'OLAP'
18DISTRIBUTED BY HASH(`id`) BUCKETS 2
19PROPERTIES (
20);

全文索引

支持对文本字段创建倒排全文索引，可实现关键词、短语检索，同时可搭配向量索引完成文本 + 向量混合检索。

SQL

1CREATE TABLE `docs` (
2  `id` int NOT NULL,
3  `title` varchar(50) NOT NULL,
4  `content` text NOT NULL,
5  INDEX idx_content (`content`) USING INVERTED PROPERTIES("parser" = "english", "lower_case" = "true", "support_phrase" = "true")
6) ENGINE=OLAP
7DUPLICATE KEY(`id`)
8DISTRIBUTED BY HASH(`id`) BUCKETS 1;
9
10CREATE TABLE `emb` (
11  `doc_id` int NOT NULL,
12  `tag` varchar(50) NOT NULL,
13  `embedding` array<float> NOT NULL,
14  INDEX idx_ann (`embedding`) USING ANN PROPERTIES("hnsw_m" = "1000", "distance" = "l2", "dimension" = "4", "algorithm" = "hnsw", "hnsw_efConstruction" = "2000")
15) ENGINE=OLAP
16DUPLICATE KEY(`doc_id`)
17DISTRIBUTED BY HASH(`doc_id`) BUCKETS 1;

DQL

下面是两张示例表的检索结果，可作为后续用例的数据参照。

SQL

1-- duplicate 表里的数据
2SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 ORDER BY __DISTANCE LIMIT 100;
3+------+-----------+------------+----------------------+------+--------------+
4| F_id | F_meta_id | F_content  | embedding            | len  | __DISTANCE   |
5+------+-----------+------------+----------------------+------+--------------+
6|    1 | meta1     | content1   | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
7|    2 | meta1     | content2   | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
8|    3 | meta2     | content1   | [1.1, 2.1, 3.1, 4.2] |    3 | 0.0099999812 |
9|    1 | meta1     | content2   | [1.1, 2.2, 3.1, 4.1] |    6 |  0.010000029 |
10|    5 | meta5     | content111 | [1.1, 2.1, 3.2, 4.2] |    6 |  0.020000011 |
11|    4 | meta3     | content3   | [2.1, 2.1, 3.1, 4.1] |    5 |   0.99999976 |
12+------+-----------+------------+----------------------+------+--------------+

SQL

1-- unique 表里的数据，F_id:6 这条是被 update 过的数据
2SELECT * FROM ann_table_unique WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 ORDER BY __DISTANCE;
3+------+-----------+-----------+----------------------+------+--------------+
4| F_id | F_meta_id | F_content | embedding            | len  | __DISTANCE   |
5+------+-----------+-----------+----------------------+------+--------------+
6|    5 | meta1     | content2  | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
7|    1 | meta1     | content1  | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
8|    2 | meta1     | content2  | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
9|    4 | meta1     | content1  | [1.1, 2.1, 3.1, 4.1] |    2 |            0 |
10|    3 | meta2     | content1  | [1.1, 2.1, 3.1, 4.2] |    3 | 0.0099999812 |
11|    6 | meta2     | content1  | [1.1, 2.1, 3.1, 4.2] |    4 | 0.0099999812 |
12+------+-----------+-----------+----------------------+------+--------------+

向量检索范围查询必须要传入期望查询返回的结果数(即limit)。如果没传，当前会设置默认值：100 范围查询没有等于。因为是近似查询。

纯向量检索

TopK 查询

标准 TopK 查询：按 ann_distance() 排序并用 LIMIT 取前 K 条。支持。

SQL

1SELECT * FROM ann_table ORDER BY ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) LIMIT 3;

image - 2026-09-11T143359.652.png

不设置 LIMIT 的 TopK 查询。支持。

SQL

1SELECT * FROM ann_table ORDER BY ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]);

image - 2026-09-11T143422.382.png

把 ann_distance() 放在 SELECT 字段里、再按别名排序。不支持。

SQL

1SELECT *, ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) AS distance FROM ann_table ORDER BY distance ASC LIMIT 5;

image - 2026-09-11T143445.233.png

范围查询

BETWEEN 范围查询 + LIMIT，不加 ORDER BY 时结果次序不固定。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 LIMIT 5;
2SELECT *, __DISTANCE FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 LIMIT 5;

image - 2026-09-11T143502.348.png

BETWEEN 范围查询不设置 LIMIT，使用默认值 100。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03;

image - 2026-09-11T143545.965.png

BETWEEN 范围查询 + 按距离全局排序，支持升序、降序和叠加 LIMIT。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 ORDER BY __DISTANCE;
2SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 ORDER BY __DISTANCE LIMIT 5;
3SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 ORDER BY __DISTANCE DESC LIMIT 5;

image - 2026-09-11T143609.282.png image - 2026-09-11T143620.754.png

用大于、小于组合出范围 + 按距离全局排序。支持。

SQL

1SELECT * FROM ann_table
2WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) > 0.01
3  AND ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) < 0.03
4ORDER BY __DISTANCE DESC LIMIT 2;

image - 2026-09-11T143642.789.png

单个距离边界查询 + 按距离全局排序，内部会自动补全另一边界的阈值。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) < 0.5 ORDER BY __DISTANCE LIMIT 10;
2SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) > 0.02 ORDER BY __DISTANCE LIMIT 10;

image - 2026-09-11T143702.124.png image - 2026-09-11T143717.385.png

查询指定列。支持。

SQL

1SELECT F_id, embedding FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 LIMIT 5;

image - 2026-09-11T143739.693.png image - 2026-09-11T143753.609.png

查询指定列并显式取出隐藏的距离列 __DISTANCE。支持。

SQL

1SELECT F_id, __DISTANCE FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 ORDER BY __DISTANCE LIMIT 5;

image - 2026-09-11T143817.847.png image - 2026-09-11T143832.569.png

子查询

派生表子查询 + 隐藏距离列。支持。

SQL

1SELECT sc.*, __DISTANCE
2FROM (SELECT embedding, len FROM ann_table_dup WHERE len > 2) AS sc
3ORDER BY ann_distance(sc.embedding, [1.1, 2.1, 3.1, 4.1])
4LIMIT 5;

image - 2026-09-11T143851.624.png

WITH 子查询 + 隐藏距离列。支持。

SQL

1WITH subquery_result AS (
2    SELECT embedding, len
3    FROM ann_table_dup
4    WHERE len > 2
5)
6SELECT *, __DISTANCE
7FROM subquery_result
8ORDER BY ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) DESC
9LIMIT 5;

image - 2026-09-11T143913.991.png

分页查询

TopK 查询 + 分页。支持。

SQL

1SELECT *, __DISTANCE FROM ann_table ORDER BY ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) LIMIT 1, 5;

image - 2026-09-11T143934.646.png

查询指定列 + 隐藏距离列 + 距离排序 + 分页。支持。

SQL

1SELECT F_id, __DISTANCE FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0.01 AND 0.03 ORDER BY __DISTANCE LIMIT 1, 5;

image - 2026-09-11T143958.833.png

复杂检索

标量过滤 + 向量检索

标量过滤 + 向量范围查询 + ORDER BY，结果顺序固定。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 AND len >= 3 ORDER BY __DISTANCE DESC;

image - 2026-09-11T144018.502.png

标量过滤 + 向量 TopK 查询。支持。

SQL

1SELECT *, __DISTANCE FROM ann_table WHERE len > 2 ORDER BY ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) LIMIT 10;

image - 2026-09-11T144035.566.png

标量过滤 + 向量范围查询，不加 ORDER BY，结果顺序不固定。支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 AND len >= 3;

image - 2026-09-11T144058.665.png

分组检索

向量范围检索 + 标量分组 + 距离聚合。支持。

SQL

1SELECT F_id, max(__DISTANCE) FROM ann_table_dup WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 GROUP BY F_id LIMIT 10;

image - 2026-09-11T144116.568.png

向量范围检索 + 标量分组 + 距离聚合 + 按聚合后的距离排序。支持。

SQL

1SELECT F_id, max(__DISTANCE) AS id_max_distance
2FROM ann_table_dup
3WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1
4GROUP BY F_id
5ORDER BY id_max_distance DESC
6LIMIT 10;

image - 2026-09-11T144135.778.png

全文检索 + 向量检索

全文检索与向量检索写在同一个 WHERE 中。不支持。

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [1.1, 2.1, 3.1, 4.1]) BETWEEN 0 AND 1 AND content MATCH_ANY 'content1';

批量向量检索

批量向量检索，即一次传入多个查询向量。不支持

SQL

1SELECT * FROM ann_table WHERE ann_distance(embedding, [[1.1, 2.1, 3.1, 4.1], [1.2, 2.1, 3.1, 4.2], [2.1, 1.1, 3.1, 4.1]]) BETWEEN 0 AND 1 AND len >= 3;

多列向量检索

用 multi_vector_search() 融合同一张表多个向量列的检索结果，融合排名结果在 __RANK 列返回。

RRF 排序

RRF排序的C参数可以不传，默认值为：60。支持。

c越小 → 高位排名加分越大（尤其是 c < 1 时，高位结果非常占优势）；

c越大 → 多个排名的权重差异变小。

SQL

1SELECT id, name, __RANK FROM multi_ann_table
2ORDER BY multi_vector_search(
3    '{"rerank":"RRF", "c":30}',
4    ann_distance(embedding1, [1.0, 2.0, 3.0, 4.0]),
5    ann_distance(embedding2, [1.1, 2.1, 3.1, 4.1])
6)
7LIMIT 5;

image - 2026-09-11T144207.622.png

WEIGHTED 排序

weight排序多列的结果。weights数组第一个值，代表后面第一列的检索结果的权重值，第二个值，代表后面第二列的检索结果的权重值。支持。

weights默认值为每列结果的权重都为1。

SQL

1SELECT id, name, __RANK FROM multi_ann_table
2ORDER BY multi_vector_search(
3    '{"rerank":"WEIGHT", "weights":[0.6,0.4]}',
4    ann_distance(embedding1, [1.0, 2.0, 3.0, 4.0]),
5    ann_distance(embedding2, [1.1, 2.1, 3.1, 4.1])
6)
7LIMIT 5;

image - 2026-09-11T144232.003.png

混合检索（多列向量 + 多列全文 + 标量过滤）

语法与多列向量检索相同，把 bm25_score() 与 ann_distance() 一起传给 hybrid_search()，标量过滤仍写在 WHERE 中。全文检索支持的 Match 类型：all、any、phrase。

RRF 排序

语法同多列向量检索。全文检索支持的Match类型：all、any、phrase。支持。

SQL

1SELECT id, name, __RANK FROM hybrid_dup_table2 WHERE age > 26 AND id > 1
2ORDER BY hybrid_search(
3    '{"rerank":"RRF"}',
4    ann_distance(embedding1, '[1.0, 2.0, 3.0, 4.0]'),
5    bm25_score(content3, 'any', 'content3'),
6    ann_distance(embedding2, '[1.1, 2.1, 3.3, 4.4]'),
7    bm25_score(content1, 'any', 'content1'),
8    bm25_score(content2, 'any', 'content2')
9) LIMIT 8;

image - 2026-09-11T144309.246.png

WEIGHTED 排序

语法同多列向量检索。全文检索支持的Match类型：all、any、phrase。支持。

SQL

1SELECT id, name, __RANK FROM hybrid_dup_table2 WHERE age > 26 AND id > 1
2ORDER BY hybrid_search(
3    '{"rerank":"WEIGHT", weights:[0.5, 0.2, 0.5, 0.2, 0.2]}',
4    ann_distance(embedding1, '[1.0, 2.0, 3.0, 4.0]'),
5    bm25_score(content3, 'any', 'content3'),
6    ann_distance(embedding2, '[1.1, 2.1, 3.3, 4.4]'),
7    bm25_score(content1, 'any', 'content1'),
8    bm25_score(content2, 'any', 'content2')
9) LIMIT 8;

image - 2026-09-11T144331.864.png

多表混合检索

用 HYBRID ... ON 连接两张表，各自召回后再融合。注意：ON 中的行键类型必须一致，例如都为 INT。

RRF 排序

docs 表全文检索 top10 与 emb 表向量检索 top10 按 RRF 融合后取 top5。支持。

要求on中的行键类型必须一致（例如都为INT）。

SQL

1SELECT __RANK,
2       COALESCE(d.id, e.doc_id) AS rid,
3       d.title,
4       d.content,
5       e.tag
6FROM docs_mt d HYBRID emb_mt e ON d.id = e.doc_id
7WHERE d.title LIKE '%apple%' AND e.tag = 'a'
8ORDER BY hybrid_search(
9    '{"rerank":"RRF","c":60,"candidates":10}',
10    bm25_score(d.content, 'any', 'apple market'),
11    ann_distance(e.embedding, '[1.0,2.0,3.0,4.0]')
12)
13LIMIT 5;

image - 2026-09-11T144357.157.png

WEIGHTED 排序

两路召回按权重融合后取 top5。支持。

SQL

1SELECT __RANK,
2       COALESCE(a.doc_id, b.doc_id) AS rid,
3       a.tag AS tag_a,
4       b.tag AS tag_b
5FROM emb_mt a HYBRID emb_mt b ON a.doc_id = b.doc_id
6WHERE a.tag IN ('a','b') AND b.doc_id <= 7
7ORDER BY hybrid_search(
8    '{"rerank":"WEIGHT","weights":[0.7,0.3]}',
9    ann_distance(a.embedding, '[1.0,2.0,3.0,4.0]'),
10    ann_distance(b.embedding, '[1.4,2.0,3.0,4.0]')
11)
12LIMIT 5;

image - 2026-09-11T144423.183.png

模型重排序（仅预览版支持）

在 hybrid_search() 中通过 reranker 指定重排序模型、query 和参与重排的文本列。

子查询中重排序

docs_mt 全文检索 top5 与 emb_mt 向量检索 top5 UNION 后，将 title、content 两列送入模型重排序。支持。

SQL

1SELECT id, title, content
2FROM (
3    -- 第一路：docs_mt 上全文召回
4    SELECT id, title, content
5    FROM (
6        SELECT id, title, content
7        FROM docs_mt
8        WHERE content MATCH_ANY 'apple market'
9        ORDER BY __SCORE DESC
10        LIMIT 5
11    ) text_recall
12    UNION ALL
13    -- 第二路：emb_mt 上向量召回
14    SELECT id, title, content
15    FROM (
16        SELECT e.doc_id AS id,
17               e.tag AS title,
18               e.tag AS content
19        FROM emb_mt e
20        WHERE e.tag IN ('a', 'b')
21        ORDER BY ann_distance(e.embedding, '[1.0,2.0,3.0,4.0]')
22        LIMIT 5
23    ) vector_recall
24) union_recall
25ORDER BY hybrid_search(
26    '{"candidates":10,
27      "reranker":{
28        "model":"jina-reranker-v2-base-multilingual",
29        "query":"apple market cap",
30        "text_columns":["title","content"]
31      }}'
32)
33LIMIT 5;

image - 2026-09-11T144501.777.png

多表 RRF 后接模型重排序

多表 RRF 融合结果再送入模型重排序。支持。

SQL

1SELECT COALESCE(d.id, e.doc_id) AS id,
2       d.title,
3       d.content,
4       e.tag
5FROM docs_mt d HYBRID emb_mt e ON d.id = e.doc_id
6ORDER BY hybrid_search(
7    '{"rerank":"RRF","c":60,"candidates":50,
8      "reranker":{
9        "model":"jina-reranker-v2-base-multilingual",
10        "query":"apple market cap",
11        "text_columns":["d.content"]
12      }}',
13    bm25_score(d.content, 'any', 'apple market'),
14    ann_distance(e.embedding, '[1.0,2.0,3.0,4.0]')
15)
16LIMIT 5;

image - 2026-09-11T144524.668.png

重排序后处理（仅预览版支持）

在 reranker 之后通过 post_process 追加后处理阶段。下例为两路检索结果依次应用 RRF 融合、模型重排序和各后处理阶段。支持。

SQL

1SELECT id, title, content, publish_time
2FROM documents
3ORDER BY hybrid_search(
4    '{
5        "rerank": "RRF",
6        "c": 60,
7        "candidates": 50,
8        "reranker": {
9            "model": "jina-reranker-v2-base-multilingual",
10            "query": "苹果公司市值",
11            "text_columns": ["title", "content"]
12        },
13        "post_process": {
14            "min_score": 0.35,
15            "length_norm": {
16                "field": "content",
17                "min_length": 50
18            },
19            "blend": {
20                "fusion_weight": 0.3,
21                "reranker_weight": 0.7
22            },
23            "decay": {
24                "function": "gauss",
25                "field": "publish_time",
26                "origin": "2026-04-20",
27                "scale": "7d",
28                "offset": "1d",
29                "decay": 0.5
30            },
31            "mmr": {
32                "lambda": 0.7,
33                "vector_column": "vec",
34                "distance_type": "cosine",
35                "top_n": 10
36            },
37            "topn_post": 10
38        }
39    }',
40    ann_distance(vec, [0.1, 0.2, 0.3, 0.4]),
41    bm25_score(content, 'any', '苹果公司市值')
42)
43LIMIT 10;

image - 2026-09-11T144553.058.png

评价此篇文章

有帮助没帮助

向量检索使用手册

ANN 算法：HNSW

百度智能云

百度胜算

百度胜算

向量检索支持SQL全集

概览

DDL

索引属性说明

Duplicate 表

Unique 表

多列索引

全文索引

DQL

纯向量检索

TopK 查询

范围查询

子查询

分页查询

复杂检索

标量过滤 + 向量检索

分组检索

全文检索 + 向量检索

批量向量检索

多列向量检索

RRF 排序

WEIGHTED 排序

混合检索（多列向量 + 多列全文 + 标量过滤）

RRF 排序

WEIGHTED 排序

多表混合检索

RRF 排序

WEIGHTED 排序

模型重排序（仅预览版支持）

子查询中重排序

多表 RRF 后接模型重排序

重排序后处理（仅预览版支持）