docs: Update notes in `tag_from_library.livemd`

public-apps/tag_from_library.livemd

@@ -5,9 +5,7 @@
 ```elixir
 Mix.install(
   [
-    # {:audio_tagger, git: "git@github.com:headwayio/audio_tagger.git"},
     {:audio_tagger, git: "https://github.com/headwayio/audio_tagger.git"},
-    # {:audio_tagger, path: "/Users/noah/development/audio_tagger"},
     {:kino_bumblebee, "~> 0.4.0"},
     {:exla, ">= 0.0.0"},
     {:explorer, "~> 0.7.0"},
@@ -26,7 +24,7 @@ audio_input = Kino.Input.audio("Audio", sampling_rate: featurizer.sampling_rate)
 ```
 
 ```elixir
-# 2 - Transcribe audio recording to text
+# 2 - Transcribe audio recording to text (using openai/whisper-tiny)
 # Takes 5–6s for about a minute of audio
 chosen_audio = Kino.Input.read(audio_input)
 file = chosen_audio.file_ref |> Kino.Input.file_path() |> File.read!()
@@ -41,11 +39,13 @@ transcription_df |> Explorer.DataFrame.head(3)
 ## Step 2: Tag Transcribed Audio
 
 ```elixir
-# Now that the audio has been transcribed to text, we run the portions of text through a
-# text-classification model to attempt to tag each with a procedure code (based on the ICD-10
-# code list).
+# Now, we run the portions of the transcribed text through the facebook/bart-large-mnli model
+# and do zero-shot text-classification to attempt to tag each with a procedure code
+# (based on the ICD-10 code list).
 
-# This takes a considerable time (1793.8s ~= 29+m) for 14 text entries with 1,000 labels.
+# The number of labels provided to the model has a direct impact on performance:
+# - 14 text entries with 1,000 labels takes a considerable time: 1793.8s total ~= 29+m
+# - 14 text entries with 10 labels produces much quicker results: 31.8s total
 
 labels_df = AudioTagger.SampleData.icd10_codes()
 tagged_audio = AudioTagger.Tagger.tag_audio(transcription_df, labels_df)