优化 Qdrant 性能：三种场景

不同的用例需要在内存使用、搜索速度和精度之间取得不同的平衡。Qdrant 的设计灵活且可定制，因此您可以根据自己的特定需求进行调整。

本指南将引导您了解三种主要的优化策略

高速搜索和低内存使用
高精度和低内存使用
高精度和高速搜索

qdrant resource tradeoffs

1. 低内存使用的高速搜索

为了在最小化内存使用的同时实现高速搜索，您可以将向量存储在磁盘上，同时最小化磁盘读取次数。向量量化是一种压缩向量的技术，允许将更多的向量存储在内存中，从而减少从磁盘读取的需要。

要配置内存量化和磁盘上的原始向量，您需要使用以下参数创建集合

on_disk：将原始向量存储在磁盘上。
quantization_config：使用 scalar 方法将量化向量压缩为 int8。
always_ram：将量化向量保留在 RAM 中。

PUT /collections/{collection_name}
{
    "vectors": {
        "size": 768,
        "distance": "Cosine",
        "on_disk": true
    },
    "quantization_config": {
        "scalar": {
            "type": "int8",
            "always_ram": true
        }
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(size=768, distance=models.Distance.COSINE, on_disk=True),
    quantization_config=models.ScalarQuantization(
        scalar=models.ScalarQuantizationConfig(
            type=models.ScalarType.INT8,
            always_ram=True,
        ),
    ),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
    on_disk: true,
  },
  quantization_config: {
    scalar: {
      type: "int8",
      always_ram: true,
    },
  },
});

use qdrant_client::qdrant::{
    CreateCollectionBuilder, Distance, QuantizationType, ScalarQuantizationBuilder,
    VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine))
            .quantization_config(
                ScalarQuantizationBuilder::default()
                    .r#type(QuantizationType::Int8.into())
                    .always_ram(true),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.OptimizersConfigDiff;
import io.qdrant.client.grpc.Collections.QuantizationConfig;
import io.qdrant.client.grpc.Collections.QuantizationType;
import io.qdrant.client.grpc.Collections.ScalarQuantization;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .setOnDisk(true)
                            .build())
                    .build())
            .setQuantizationConfig(
                QuantizationConfig.newBuilder()
                    .setScalar(
                        ScalarQuantization.newBuilder()
                            .setType(QuantizationType.Int8)
                            .setAlwaysRam(true)
                            .build())
                    .build())
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine, OnDisk = true },
	quantizationConfig: new QuantizationConfig
	{
		Scalar = new ScalarQuantization { Type = QuantizationType.Int8, AlwaysRam = true }
	}
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
		OnDisk:   qdrant.PtrOf(true),
	}),
	QuantizationConfig: qdrant.NewQuantizationScalar(&qdrant.ScalarQuantization{
		Type:      qdrant.QuantizationType_Int8,
		AlwaysRam: qdrant.PtrOf(true),
	}),
})

禁用重新评分以加快搜索速度（可选）

这是完全可选的。使用搜索 params 禁用重新评分可以进一步减少磁盘读取次数。请注意，这可能会略微降低精度。

POST /collections/{collection_name}/points/query
{
    "query": [0.2, 0.1, 0.9, 0.7],
    "params": {
        "quantization": {
            "rescore": false
        }
    },
    "limit": 10
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.query_points(
    collection_name="{collection_name}",
    query=[0.2, 0.1, 0.9, 0.7],
    search_params=models.SearchParams(
        quantization=models.QuantizationSearchParams(rescore=False)
    ),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.query("{collection_name}", {
    query: [0.2, 0.1, 0.9, 0.7],
    params: {
        quantization: {
            rescore: false,
        },
    },
});

use qdrant_client::qdrant::{
    QuantizationSearchParamsBuilder, QueryPointsBuilder, SearchParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .query(
        QueryPointsBuilder::new("{collection_name}")
            .query(vec![0.2, 0.1, 0.9, 0.7])
            .limit(3)
            .params(
                SearchParamsBuilder::default()
                    .quantization(QuantizationSearchParamsBuilder::default().rescore(false)),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Points.QuantizationSearchParams;
import io.qdrant.client.grpc.Points.QueryPoints;
import io.qdrant.client.grpc.Points.SearchParams;

import static io.qdrant.client.QueryFactory.nearest;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client.queryAsync(
        QueryPoints.newBuilder()
                .setCollectionName("{collection_name}")
                .setQuery(nearest(0.2f, 0.1f, 0.9f, 0.7f))
                .setParams(
                        SearchParams.newBuilder()
                                .setQuantization(
                                        QuantizationSearchParams.newBuilder().setRescore(false).build())
                                .build())
                .setLimit(3)
                .build())
        .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.QueryAsync(
	collectionName: "{collection_name}",
	query: new float[] { 0.2f, 0.1f, 0.9f, 0.7f },
	searchParams: new SearchParams
	{
		Quantization = new QuantizationSearchParams { Rescore = false }
	},
	limit: 3
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.Query(context.Background(), &qdrant.QueryPoints{
	CollectionName: "{collection_name}",
	Query:          qdrant.NewQuery(0.2, 0.1, 0.9, 0.7),
	Params: &qdrant.SearchParams{
		Quantization: &qdrant.QuantizationSearchParams{
			Rescore: qdrant.PtrOf(true),
		},
	},
})

2. 低内存使用的高精度

如果您需要高精度但 RAM 有限，可以将向量和 HNSW 索引都存储在磁盘上。此设置可减少内存使用，同时保持搜索精度。

要在 on_disk 上存储向量，您需要同时配置向量和 HNSW 索引

PUT /collections/{collection_name}
{
    "vectors": {
      "size": 768,
      "distance": "Cosine",
      "on_disk": true
    },
    "hnsw_config": {
        "on_disk": true
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(size=768, distance=models.Distance.COSINE, on_disk=True),
    hnsw_config=models.HnswConfigDiff(on_disk=True),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
    on_disk: true,
  },
  hnsw_config: {
    on_disk: true,
  },
});

use qdrant_client::qdrant::{
    CreateCollectionBuilder, Distance, HnswConfigDiffBuilder,
    VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine).on_disk(true))
            .hnsw_config(HnswConfigDiffBuilder::default().on_disk(true)),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.HnswConfigDiff;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .setOnDisk(true)
                            .build())
                    .build())
            .setHnswConfig(HnswConfigDiff.newBuilder().setOnDisk(true).build())
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine, OnDisk = true },
	hnswConfig: new HnswConfigDiff { OnDisk = true }
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
		OnDisk:   qdrant.PtrOf(true),
	}),
	HnswConfig: &qdrant.HnswConfigDiff{
		OnDisk: qdrant.PtrOf(true),
	},
})

提高精度

即使在 RAM 有限的情况下，也要增加 HNSW 索引的 ef 和 m 参数以提高精度

...
"hnsw_config": {
    "m": 64,
    "ef_construct": 512,
    "on_disk": true
}
...

注意：此设置的速度取决于磁盘的 IOPS（每秒输入/输出操作）。
您可以使用 fio 来测量磁盘 IOPS。

HNSW 索引中的内联存储

自 v1.16.0 起可用

当在磁盘上存储向量和 HNSW 索引时，您可以通过在 hnsw_config 中启用 inline_storage 选项来提高搜索性能。启用内联存储后，Qdrant 会将向量副本直接存储在 HNSW 索引文件中。这通过减少 IO 操作数量来加快搜索速度，但会增加 3-4 倍的存储使用量。它需要启用量化。

PUT /collections/{collection_name}
{
    "vectors": {
      "size": 768,
      "distance": "Cosine",
      "on_disk": true
    },
    "quantization_config": {
        "binary": {
            "always_ram": false
        }
    },
    "hnsw_config": {
        "on_disk": true,
        "inline_storage": true
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(
        size=768, distance=models.Distance.COSINE, on_disk=True
    ),
    quantization_config=models.BinaryQuantization(
        binary=models.BinaryQuantizationConfig(always_ram=False),
    ),
    hnsw_config=models.HnswConfigDiff(on_disk=True, inline_storage=True),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
    on_disk: true,
  },
  quantization_config: {
    binary: {
      always_ram: false,
    },
  },
  hnsw_config: {
    on_disk: true,
    inline_storage: true,
  },
});

use qdrant_client::qdrant::{
    BinaryQuantizationBuilder, CreateCollectionBuilder, Distance, HnswConfigDiffBuilder,
    VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine).on_disk(true))
            .quantization_config(BinaryQuantizationBuilder::new(false))
            .hnsw_config(
                HnswConfigDiffBuilder::default()
                    .on_disk(true)
                    .inline_storage(true),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.BinaryQuantization;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.HnswConfigDiff;
import io.qdrant.client.grpc.Collections.QuantizationConfig;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .setOnDisk(true)
                            .build())
                    .build())
            .setQuantizationConfig(
                QuantizationConfig.newBuilder()
                    .setBinary(BinaryQuantization.newBuilder().setAlwaysRam(false).build())
                    .build())
            .setHnswConfig(HnswConfigDiff.newBuilder().setOnDisk(true).setInlineStorage(true).build())
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine, OnDisk = true },
	quantizationConfig: new QuantizationConfig
	{
		Binary = new BinaryQuantization { AlwaysRam = false }
	},
	hnswConfig: new HnswConfigDiff { OnDisk = true, InlineStorage = true }
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
		OnDisk:   qdrant.PtrOf(true),
	}),
	QuantizationConfig: qdrant.NewQuantizationBinary(
		&qdrant.BinaryQuantization{
			AlwaysRam: qdrant.PtrOf(false),
		},
	),
	HnswConfig: &qdrant.HnswConfigDiff{
		OnDisk:        qdrant.PtrOf(true),
		InlineStorage: qdrant.PtrOf(true),
	},
})

3. 高精度和高速搜索

对于需要高速和高精度的场景，请尽可能多地将数据保留在 RAM 中。使用带有重新评分的量化以实现可调精度。

以下是如何为集合配置标量量化

PUT /collections/{collection_name}
{
    "vectors": {
      "size": 768,
      "distance": "Cosine"
    },
    "quantization_config": {
        "scalar": {
            "type": "int8",
            "always_ram": true
        }
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(size=768, distance=models.Distance.COSINE),
    quantization_config=models.ScalarQuantization(
        scalar=models.ScalarQuantizationConfig(
            type=models.ScalarType.INT8,
            always_ram=True,
        ),
    ),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
  },
  quantization_config: {
    scalar: {
      type: "int8",
      always_ram: true,
    },
  },
});

use qdrant_client::qdrant::{
    CreateCollectionBuilder, Distance, QuantizationType, ScalarQuantizationBuilder,
    VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine))
            .quantization_config(
                ScalarQuantizationBuilder::default()
                    .r#type(QuantizationType::Int8.into())
                    .always_ram(true),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.OptimizersConfigDiff;
import io.qdrant.client.grpc.Collections.QuantizationConfig;
import io.qdrant.client.grpc.Collections.QuantizationType;
import io.qdrant.client.grpc.Collections.ScalarQuantization;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .build())
                    .build())
            .setQuantizationConfig(
                QuantizationConfig.newBuilder()
                    .setScalar(
                        ScalarQuantization.newBuilder()
                            .setType(QuantizationType.Int8)
                            .setAlwaysRam(true)
                            .build())
                    .build())
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine},
	quantizationConfig: new QuantizationConfig
	{
		Scalar = new ScalarQuantization { Type = QuantizationType.Int8, AlwaysRam = true }
	}
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
	}),
	QuantizationConfig: qdrant.NewQuantizationScalar(&qdrant.ScalarQuantization{
		Type:      qdrant.QuantizationType_Int8,
		AlwaysRam: qdrant.PtrOf(true),
	}),
})

微调搜索参数

您可以调整 hnsw_ef 和 exact 等搜索参数，以在速度和精度之间取得平衡

关键参数

hnsw_ef：搜索期间访问的邻居数量（值越高 = 精度越高，速度越慢）。
exact：设置为 true 可进行精确搜索，这会更慢但更准确。您可以使用它来比较使用不同 hnsw_ef 值进行搜索的结果与真实情况。

POST /collections/{collection_name}/points/query
{
    "query": [0.2, 0.1, 0.9, 0.7],
    "params": {
        "hnsw_ef": 128,
        "exact": false
    },
    "limit": 3
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.query_points(
    collection_name="{collection_name}",
    query=[0.2, 0.1, 0.9, 0.7],
    search_params=models.SearchParams(hnsw_ef=128, exact=False),
    limit=3,
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.query("{collection_name}", {
    query: [0.2, 0.1, 0.9, 0.7],
    params: {
        hnsw_ef: 128,
        exact: false,
    },
    limit: 3,
});

use qdrant_client::qdrant::{QueryPointsBuilder, SearchParamsBuilder};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .query(
        QueryPointsBuilder::new("{collection_name}")
            .query(vec![0.2, 0.1, 0.9, 0.7])
            .limit(3)
            .params(SearchParamsBuilder::default().hnsw_ef(128).exact(false)),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Points.QueryPoints;
import io.qdrant.client.grpc.Points.SearchParams;

import static io.qdrant.client.QueryFactory.nearest;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client.queryAsync(
        QueryPoints.newBuilder()
                .setCollectionName("{collection_name}")
                .setQuery(nearest(0.2f, 0.1f, 0.9f, 0.7f))
                .setParams(SearchParams.newBuilder().setHnswEf(128).setExact(false).build())
                .setLimit(3)
                .build())
        .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.QueryAsync(
	collectionName: "{collection_name}",
	query: new float[] { 0.2f, 0.1f, 0.9f, 0.7f },
	searchParams: new SearchParams { HnswEf = 128, Exact = false },
	limit: 3
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.Query(context.Background(), &qdrant.QueryPoints{
	CollectionName: "{collection_name}",
	Query:          qdrant.NewQuery(0.2, 0.1, 0.9, 0.7),
	Params: &qdrant.SearchParams{
		HnswEf: qdrant.PtrOf(uint64(128)),
		Exact:  qdrant.PtrOf(false),
	},
})

平衡延迟和吞吐量

在优化搜索性能时，延迟和吞吐量是两个需要考虑的主要指标

延迟：单个请求所需的时间。
吞吐量：每秒处理的请求数量。

以下优化方法并非相互排斥，但在某些情况下，可能更倾向于优化其中一个。

最小化延迟

为了最小化延迟，您可以将 Qdrant 设置为对单个请求使用尽可能多的核心。您可以通过将集合中的段数设置为等于系统中的核心数来实现此目的。

在这种情况下，每个段将并行处理，并且最终结果将更快获得。

PUT /collections/{collection_name}
{
    "vectors": {
      "size": 768,
      "distance": "Cosine"
    },
    "optimizers_config": {
        "default_segment_number": 16
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(size=768, distance=models.Distance.COSINE),
    optimizers_config=models.OptimizersConfigDiff(default_segment_number=16),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
  },
  optimizers_config: {
    default_segment_number: 16,
  },
});

use qdrant_client::qdrant::{
    CreateCollectionBuilder, Distance, OptimizersConfigDiffBuilder, VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine))
            .optimizers_config(
                OptimizersConfigDiffBuilder::default().default_segment_number(16),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.OptimizersConfigDiff;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .build())
                    .build())
            .setOptimizersConfig(
                OptimizersConfigDiff.newBuilder().setDefaultSegmentNumber(16).build())
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine },
	optimizersConfig: new OptimizersConfigDiff { DefaultSegmentNumber = 16 }
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
	}),
	OptimizersConfig: &qdrant.OptimizersConfigDiff{
		DefaultSegmentNumber: qdrant.PtrOf(uint64(16)),
	},
})

最大化吞吐量

为了最大化吞吐量，请将 Qdrant 配置为使用尽可能多的核心并行处理多个请求。

为此，请使用更少的段（通常为 2）和更大的大小（默认每个段 200MB）来并行处理更多请求。

大型段受益于索引的大小和查找最近邻居所需的整体较少的向量比较。但是，它们需要更多时间来构建 HNSW 索引。

PUT /collections/{collection_name}
{
    "vectors": {
      "size": 768,
      "distance": "Cosine"
    },
    "optimizers_config": {
        "default_segment_number": 2,
        "max_segment_size": 5000000
    }
}

from qdrant_client import QdrantClient, models

client = QdrantClient(url="https://:6333")

client.create_collection(
    collection_name="{collection_name}",
    vectors_config=models.VectorParams(size=768, distance=models.Distance.COSINE),
    optimizers_config=models.OptimizersConfigDiff(default_segment_number=2, max_segment_size=5000000),
)

import { QdrantClient } from "@qdrant/js-client-rest";

const client = new QdrantClient({ host: "localhost", port: 6333 });

client.createCollection("{collection_name}", {
  vectors: {
    size: 768,
    distance: "Cosine",
  },
  optimizers_config: {
    default_segment_number: 2,
    max_segment_size: 5000000,
  },
});

use qdrant_client::qdrant::{
    CreateCollectionBuilder, Distance, OptimizersConfigDiffBuilder, VectorParamsBuilder,
};
use qdrant_client::Qdrant;

let client = Qdrant::from_url("https://:6334").build()?;

client
    .create_collection(
        CreateCollectionBuilder::new("{collection_name}")
            .vectors_config(VectorParamsBuilder::new(768, Distance::Cosine))
            .optimizers_config(
                OptimizersConfigDiffBuilder::default().default_segment_number(2).max_segment_size(5000000),
            ),
    )
    .await?;

import io.qdrant.client.QdrantClient;
import io.qdrant.client.QdrantGrpcClient;
import io.qdrant.client.grpc.Collections.CreateCollection;
import io.qdrant.client.grpc.Collections.Distance;
import io.qdrant.client.grpc.Collections.OptimizersConfigDiff;
import io.qdrant.client.grpc.Collections.VectorParams;
import io.qdrant.client.grpc.Collections.VectorsConfig;

QdrantClient client =
    new QdrantClient(QdrantGrpcClient.newBuilder("localhost", 6334, false).build());

client
    .createCollectionAsync(
        CreateCollection.newBuilder()
            .setCollectionName("{collection_name}")
            .setVectorsConfig(
                VectorsConfig.newBuilder()
                    .setParams(
                        VectorParams.newBuilder()
                            .setSize(768)
                            .setDistance(Distance.Cosine)
                            .build())
                    .build())
            .setOptimizersConfig(
                OptimizersConfigDiff.newBuilder()
                    .setDefaultSegmentNumber(2)
                    .setMaxSegmentSize(5000000)
                    .build()
            )
            .build())
    .get();

using Qdrant.Client;
using Qdrant.Client.Grpc;

var client = new QdrantClient("localhost", 6334);

await client.CreateCollectionAsync(
	collectionName: "{collection_name}",
	vectorsConfig: new VectorParams { Size = 768, Distance = Distance.Cosine },
	optimizersConfig: new OptimizersConfigDiff { DefaultSegmentNumber = 2, MaxSegmentSize = 5000000 }
);

import (
	"context"

	"github.com/qdrant/go-client/qdrant"
)

client, err := qdrant.NewClient(&qdrant.Config{
	Host: "localhost",
	Port: 6334,
})

client.CreateCollection(context.Background(), &qdrant.CreateCollection{
	CollectionName: "{collection_name}",
	VectorsConfig: qdrant.NewVectorsConfig(&qdrant.VectorParams{
		Size:     768,
		Distance: qdrant.Distance_Cosine,
	}),
	OptimizersConfig: &qdrant.OptimizersConfigDiff{
		DefaultSegmentNumber: qdrant.PtrOf(uint64(2)),
		MaxSegmentSize:       qdrant.PtrOf(uint64(5000000)),
	},
})

总结

通过调整向量存储、量化和搜索参数等配置，您可以针对不同的用例优化 Qdrant

低内存 + 高速：使用向量量化。
高精度 + 低内存：将向量和 HNSW 索引存储在磁盘上。
高精度 + 高速：将数据保留在 RAM 中，使用带有重新评分的量化。
延迟 vs. 吞吐量：根据优先级调整段数。

选择最适合您用例的策略，以最大限度地发挥 Qdrant 的性能。