HNSW

这是 hnswlib 的一个包装器，用于将 CAGRA 索引加载为不可变的 HNSW 索引。加载的 HNSW 索引仅在 cuVS 中兼容，可以使用包装函数进行搜索。

#include <raft/neighbors/hnsw.h>

索引搜索参数

typedef struct cuvsHnswSearchParams *cuvsHnswSearchParams_t

cuvsError_t cuvsHnswSearchParamsCreate(

cuvsHnswSearchParams_t *params

)

分配 HNSW 搜索参数，并填充默认值。

参数:

params – [in] 要分配的 cuvsHnswSearchParams_t

返回:

cuvsError_t

cuvsError_t cuvsHnswSearchParamsDestroy(

cuvsHnswSearchParams_t params

)

释放 HNSW 搜索参数。

参数:

params – [in] 要释放的 cuvsHnswSearchParams_t

返回:

cuvsError_t

struct cuvsHnswSearchParams

#include <hnsw.h>

索引

typedef cuvsHnswIndex *cuvsHnswIndex_t

cuvsError_t cuvsHnswIndexCreate(cuvsHnswIndex_t *index)

分配 HNSW 索引。

参数:

index – [in] 要分配的 cuvsHnswIndex_t

返回:

HnswError_t

cuvsError_t cuvsHnswIndexDestroy(cuvsHnswIndex_t index)

释放 HNSW 索引。

参数:

index – [in] 要释放的 cuvsHnswIndex_t

struct cuvsHnswIndex

#include <hnsw.h>

结构体，用于保存 cuvs::neighbors::Hnsw::index 的地址及其活跃的训练 dtype。

索引扩展参数

typedef struct cuvsHnswExtendParams *cuvsHnswExtendParams_t

cuvsError_t cuvsHnswExtendParamsCreate(

cuvsHnswExtendParams_t *params

)

分配 HNSW 扩展参数，并填充默认值。

参数:

params – [in] 要分配的 cuvsHnswExtendParams_t

返回:

cuvsError_t

cuvsError_t cuvsHnswExtendParamsDestroy(

cuvsHnswExtendParams_t params

)

释放 HNSW 扩展参数。

参数:

params – [in] 要释放的 cuvsHnswExtendParams_t

返回:

cuvsError_t

struct cuvsHnswExtendParams

#include <hnsw.h>

公共成员

int num_threads

用于扩展附加向量的 CPU 线程数

索引扩展

cuvsError_t cuvsHnswExtend(

cuvsResources_t res,

cuvsHnswExtendParams_t params,

cuvsHnswExtendParams_t params,

DLManagedTensor *additional_dataset

)

cuvsHnswIndex_t index

#include <cuvs/core/c_api.h>
#include <cuvs/neighbors/cagra.h>
#include <cuvs/neighbors/hnsw.h>

// Create cuvsResources_t
cuvsResources_t res;
cuvsError_t res_create_status = cuvsResourcesCreate(&res);

// create an index with `cuvsCagraBuild`

// Convert the CAGRA index to an HNSW index
cuvsHnswIndex_t hnsw_index;
cuvsHnswIndexCreate(&hnsw_index);
cuvsHnswIndexParams_t hnsw_params;
cuvsHnswIndexParamsCreate(&hnsw_params);
cuvsHnswFromCagra(res, hnsw_params, cagra_index, hnsw_index);

// Extend the HNSW index with additional vectors
DLManagedTensor additional_dataset;
cuvsHnswExtendParams_t extend_params;
cuvsHnswExtendParamsCreate(&extend_params);
cuvsHnswExtend(res, extend_params, additional_dataset, hnsw_index);

// de-allocate `hnsw_params`, `hnsw_index`, `extend_params` and `res`
cuvsError_t hnsw_params_destroy_status = cuvsHnswIndexParamsDestroy(hnsw_params);
cuvsError_t hnsw_index_destroy_status = cuvsHnswIndexDestroy(hnsw_index);
cuvsError_t extend_params_destroy_status = cuvsHnswExtendParamsDestroy(extend_params);
cuvsError_t res_destroy_status = cuvsResourcesDestroy(res);

参数:

• 将新向量添加到 HNSW 索引 注意：只有当层级为 CPU 时，HNSW 索引才能在从 CAGRA 索引转换后进行扩展。

• res – [in] cuvsResources_t 不透明 C 句柄

• params – [in] 用于扩展 Hnsw 索引的 cuvsHnswExtendParams_t

• additional_dataset – [in] DLManagedTensor* 附加数据集，用于扩展索引

返回:

cuvsError_t

index – [inout] 要扩展的 cuvsHnswIndex_t

索引加载

cuvsResources_t res,

cuvsError_t cuvsHnswFromCagra(,

cuvsResources_t res,

cuvsHnswIndexParams_t params

)

cuvsCagraIndex_t cagra_index

1. cuvsHnswIndex_t hnsw_index

2. 将 CAGRA 索引转换为 HNSW 索引。注意：当层级为

#include <cuvs/core/c_api.h>
#include <cuvs/neighbors/cagra.h>
#include <cuvs/neighbors/hnsw.h>

// Create cuvsResources_t
cuvsResources_t res;
cuvsError_t res_create_status = cuvsResourcesCreate(&res);

// create a CAGRA index with `cuvsCagraBuild`

// Convert the CAGRA index to an HNSW index
cuvsHnswIndex_t hnsw_index;
cuvsHnswIndexCreate(&hnsw_index);
cuvsHnswIndexParams_t hnsw_params;
cuvsHnswIndexParamsCreate(&hnsw_params);
cuvsHnswFromCagra(res, hnsw_params, cagra_index, hnsw_index);

// de-allocate `hnsw_params`, `hnsw_index` and `res`
cuvsError_t hnsw_params_destroy_status = cuvsHnswIndexParamsDestroy(hnsw_params);
cuvsError_t hnsw_index_destroy_status = cuvsHnswIndexDestroy(hnsw_index);
cuvsError_t res_destroy_status = cuvsResourcesDestroy(res);

参数:

• 将新向量添加到 HNSW 索引 注意：只有当层级为 CPU 时，HNSW 索引才能在从 CAGRA 索引转换后进行扩展。

• NONE: 此方法使用文件系统将 CAGRA 索引写入 /tmp/<random_number>.bin，然后将其读取为 hnswlib 索引，最后删除临时文件。返回的索引是不可变的，只能由 cuVS 中的 hnswlib 包装器进行搜索，因为格式与原始 hnswlib 不兼容。

• CPU: 返回的索引是可变的，可以使用附加向量进行扩展。序列化后的索引也与原始 hnswlib 库兼容。

• res – [in] cuvsResources_t 不透明 C 句柄

返回:

cuvsError_t

params – [in] 用于加载 Hnsw 索引的 cuvsHnswIndexParams_t

cuvsResources_t res,

cuvsError_t cuvsHnswFromCagra(,

cuvsResources_t res,

cuvsHnswIndexParams_t params,

cagra_index – [in] 要转换为 HNSW 索引的 cuvsCagraIndex_t

)

hnsw_index – [out] 用于返回 HNSW 索引的 cuvsHnswIndex_t

cuvsError_t cuvsHnswFromCagraWithDataset(

cuvsResources_t res,

cuvsHnswSearchParams_t params,

DLManagedTensor *additional_dataset,

cuvsResources_t res,

cuvsHnswIndexParams_t params,

cuvsCagraIndex_t cagra_index

)

cuvsHnswIndex_t hnsw_index

1. DLManagedTensor *dataset_tensor

2. 索引搜索

3. cuvsError_t cuvsHnswSearch(

#include <cuvs/core/c_api.h>
#include <cuvs/neighbors/hnsw.h>

// Create cuvsResources_t
cuvsResources_t res;
cuvsError_t res_create_status = cuvsResourcesCreate(&res);

// Assume a populated `DLManagedTensor` type here
DLManagedTensor dataset;
DLManagedTensor queries;
DLManagedTensor neighbors;

// Create default search params
cuvsHnswSearchParams_t params;
cuvsError_t params_create_status = cuvsHnswSearchParamsCreate(&params);

// Search the `index` built using `cuvsHnswFromCagra`
cuvsError_t search_status = cuvsHnswSearch(res, params, index, &queries, &neighbors,
&distances);

// de-allocate `params` and `res`
cuvsError_t params_destroy_status = cuvsHnswSearchParamsDestroy(params);
cuvsError_t res_destroy_status = cuvsResourcesDestroy(res);

参数:

• 将新向量添加到 HNSW 索引 注意：只有当层级为 CPU 时，HNSW 索引才能在从 CAGRA 索引转换后进行扩展。

• cuvsResources_t res

• cuvsHnswSearchParams_t params

• cuvsHnswIndex_t index

• DLManagedTensor *queries

• DLManagedTensor *neighbors

DLManagedTensor *distances

使用底层 DLDeviceType 等于 kDLCPU、kDLCUDAHost 或 kDLCUDAManaged 的 DLManagedTensor 搜索 HNSW 索引。同样重要的是要注意，HNSW 索引必须使用与 queries 相同的数据类型构建，即 index.dtype.code == queries.dl_tensor.dtype.code 支持的输入类型有

queries: a. kDLDataType.code == kDLFloat 且 kDLDataType.bits = 32 b. kDLDataType.code == kDLInt 且 kDLDataType.bits = 8 c. kDLDataType.code == kDLUInt 且 kDLDataType.bits = 8