Class: Candle::Reranker

Inherits:
Object
  • Object
show all
Defined in:
lib/candle/reranker.rb

Constant Summary collapse

DEFAULT_MODEL_PATH =

Default model path for cross-encoder/ms-marco-MiniLM-L-12-v2

"cross-encoder/ms-marco-MiniLM-L-12-v2"

Class Method Summary collapse

Instance Method Summary collapse

Class Method Details

.new(model_path: DEFAULT_MODEL_PATH, device: Candle::Device.cpu) ⇒ Object

Constructor for creating a new Reranker with optional parameters

Parameters:

  • model_path (String, nil) (defaults to: DEFAULT_MODEL_PATH) —

    The path to the model on Hugging Face

  • device (Candle::Device, Candle::Device.cpu) (defaults to: Candle::Device.cpu) —

    The device to use for computation



9
10
11
 
# File 'lib/candle/reranker.rb', line 9

def self.new(model_path: DEFAULT_MODEL_PATH, device: Candle::Device.cpu)
  _create(model_path, device)
end

Instance Method Details

#rerank(query, documents, pooling_method: "pooler", apply_sigmoid: true) ⇒ Object

Returns documents ranked by relevance using the specified pooling method.

Parameters:

  • query (String) —

    The input text

  • documents (Array<String>) —

    The list of documents to compare against

  • pooling_method (String) (defaults to: "pooler") —

    Pooling method: “pooler”, “cls”, or “mean”. Default: “pooler”

  • apply_sigmoid (Boolean) (defaults to: true) —

    Whether to apply sigmoid to the scores. Default: true



18
19
20
21
22
 
# File 'lib/candle/reranker.rb', line 18

def rerank(query, documents, pooling_method: "pooler", apply_sigmoid: true)
  rerank_with_options(query, documents, pooling_method, apply_sigmoid).collect { |doc, score, doc_id|
    { doc_id: doc_id, score: score, text: doc }
  }
end