Approximate Retrieval

AscendIndexIVFSP

• Feature type: FP32
• Feature dimension: 64, 128, 256, 512, or 768
• Distance type: L2
• Calculation precision: medium
• Device memory usage: low (compressed features)
• Application scenario: approximate retrieval in a large-scale database (hundreds of millions of records) with high performance requirements and tolerance to precision loss
• It is supported only by the Atlas inference product.

• IVFSP service operator
• IVFSP AICPU operator
• IVFSP training operator (used only for codebook training)

For details, see IVFSP.

Link

AscendIndexIVFSQ

• Feature type: FP32
• Feature dimension: 64, 128, 256, 384, or 512
• Distance type: L2 or IP
• Calculation precision: medium
• Device memory usage: low (quantized to int8)
• Application scenario: IVFSQ used as a performance-precision regulator in scenarios with high requirements on performance and tolerance to precision loss

• IVFSQ8
• AICPU
• FlatAT (The FlatAT operator needs to be generated only when useKmeansPP is set to true.)

Link

AscendIndexIVFSQT

• Feature type: FP32
• Feature dimension: 256
• Distance type: IP
• Calculation precision: medium
• Device memory usage: low (quantization and dimension reduction)
• Application scenario: AscendIndexIVFSQT is an IVFSQ algorithm for three-level retrieval and contains a dimension reduction algorithm. It is applicable to approximate retrieval in a large-scale database (hundreds of millions of records) with high performance requirements and tolerance to precision loss.

• IVFSQT
• FlatAT
• AICPU
• FlatInt8AT (required on the Atlas inference product)

Link

AscendIndexBinaryFlat

• Feature type: uint8 binarization features
• Feature dimension: 256, 512, or 1024
• Distance type: Hamming and IP
• Calculation precision: high
• Device memory usage: low
• Application scenario: The AscendIndexBinaryFlat class inherits from the Faiss IndexBinary class and is used for binarized feature retrieval. It is used in scenarios requiring low memory usage and high performance.
• It is supported only by the Atlas inference product.

• BinaryFlat
• AICPU

Link

AscendIndexVStar

• Feature type: FP32
• Feature dimension: 128, 256, 512, or 1024
• Distance type: L2
• Calculation precision: medium
• Device memory usage: low (compressed features)
• Application scenario: approximate retrieval in a large-scale database (tens of millions of records) with high performance requirements and tolerance to precision loss
• It is supported only by the Atlas inference product.

• VStar service operator
• VStar AICPU Operator
• VStar training operator (used only for codebook training)

  For details, see VSTAR.

Link

AscendIndexGreat

• Feature type: FP32
• Feature dimension: 128, 256, 512, or 1024
• Distance type: L2
• Calculation precision: medium
• Device memory usage: low (compressed features)
• Application scenario: approximate retrieval in a large-scale database (tens of millions of records) with high performance requirements and tolerance to precision loss
• It is supported only by the Atlas inference product.

(An operator needs to be generated only when mode is set to AKMode.)

• VStar service operator
• VStar AICPU Operator
• VStar training operator (used only for codebook training)

For details, see VSTAR.

Link

AscendIndexIVFFlat

• Feature type: FP32
• Feature dimension: 128
• Distance type: IP
• Calculation precision: medium
• Device memory usage: low (compressed features)
• Application scenario: approximate retrieval in a large-scale database (hundreds of millions of records) with high performance requirements and tolerance to precision loss
• It is supported by the Atlas A2 inference products and Atlas A3 inference products.

• AICPU
• IVFFLAT

Link