lib/medical_transcription/code_searcher.ex

defmodule MedicalTranscription.CodeSearcher do
  @moduledoc """
  Takes a portion of text and searches for closely matching results within a list of vectors, by using AudioTagger's
  SemanticSearch module.
  """

  @input_filename "icd9_codelist"

  alias AudioTagger.Classifier.SemanticSearch
  alias AudioTagger.Structs.SemanticSearchConfiguration

  def prepare_search_configuration() do
    {model_info, tokenizer} = SemanticSearch.prepare_model()
    labels_df = read_labels_from_csv!()

    label_embeddings = SemanticSearch.load_label_vectors(vectors_filepath())

    %SemanticSearchConfiguration{
      labels_df: labels_df,
      label_embeddings: label_embeddings,
      model_info: model_info,
      tokenizer: tokenizer,
      opts: [similarity_threshold: 0.8]
    }
  end

  def process_chunk(%SemanticSearchConfiguration{} = input, text) do
    SemanticSearch.tag_one(input, text)
  end

  defp read_labels_from_csv! do
    Explorer.DataFrame.from_csv!(
      labels_filepath(),
      dtypes: [
        {"code", :string},
        {"long_description", :string}
      ]
    )
  end

  defp input_filepath() do
    Path.join(AudioTagger.SampleData.cache_dir(), @input_filename)
  end

  defp vectors_filepath(), do: "#{input_filepath()}.bin"
  defp labels_filepath(), do: "#{input_filepath()}.csv"
end