Skip to main content

Interface: LocalOnnxEmbedderConfig

Defined in: Services/Search/SearchPipelineConfig.ts:24

Properties​

dimensions?​

optional dimensions?: number

Defined in: Services/Search/SearchPipelineConfig.ts:29

Embedding dimensions (for Matryoshka truncation). Default: 768

dtype?​

optional dtype?: string

Defined in: Services/Search/SearchPipelineConfig.ts:46

ONNX model quantization dtype. Default: 'q8'.

Measured for nomic-embed-text-v1.5: 'q8' is ~317 MB resident and ~13 ms per embed; 'fp32' is ~1,053 MB and ~28 ms. q8 is the default because the index it is searched against is itself int8-quantized, so fp32 query precision buys very little. Set 'fp32' to trade memory for maximum fidelity.

The chosen dtype must exist in the model directory — a self-hosted model with allowRemoteModels: false will fail to load if the matching ONNX file was never downloaded.

meanVectorPath?​

optional meanVectorPath?: string

Defined in: Services/Search/SearchPipelineConfig.ts:33

Path to mean vector JSON file for centering (anisotropy correction). If set, centering is applied after truncation.

modelName?​

optional modelName?: string

Defined in: Services/Search/SearchPipelineConfig.ts:27

HuggingFace model name. Default: 'nomic-ai/nomic-embed-text-v1.5'

provider​

provider: "local-onnx"

Defined in: Services/Search/SearchPipelineConfig.ts:25

queryPrefix?​

optional queryPrefix?: string

Defined in: Services/Search/SearchPipelineConfig.ts:31

Prefix prepended to query text. Default: 'search_query: '