Update song-insight-app.py

song-insight-app.py

@@ -6,6 +6,7 @@ from langchain.chains import LLMChain
 from langchain_community.retrievers import WikipediaRetriever
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from langchain_google_genai import ChatGoogleGenerativeAI
+from google.generativeai.types.safety_types import HarmBlockThreshold, HarmCategory
 import os
 
 
@@ -18,12 +19,15 @@ def song_insight(song, artist):
     docs = retriever.get_relevant_documents(query=query_input)
 
     # LLM model
-    # llm = ChatOpenAI(openai_api_key=os.environ['OPENAI_API_KEY'], model_name="gpt-3.5-turbo", temperature=0)
-    llm = ChatGoogleGenerativeAI(model="gemini-pro", google_api_key=os.environ['GOOGLE_API_KEY'])
-
-    # Emotion Classifier Model
-    tokenizer = AutoTokenizer.from_pretrained("yangswei/emotion_text_classification")
-    emotion_model = AutoModelForSequenceClassification.from_pretrained("yangswei/emotion_text_classification")
+    # llm = ChatOpenAI(openai_api_key=OPENAI_API_KEY, model_name="gpt-3.5-turbo", temperature=0)
+    safety_setting = {
+      HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE,
+      HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_NONE,
+      HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
+      HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE,
+      }
+    llm = ChatGoogleGenerativeAI(model="gemini-pro", google_api_key=GOOGLE_API_KEY, temperature=0,
+                                 safety_settings=safety_setting, top_p=0)
 
     # get the song meaning
     template_song_meaning = """
@@ -31,43 +35,42 @@ def song_insight(song, artist):
 
       {content}
 
-      based on the the content above what does the song {song} by {artist} tell us about? give me a long explanations
+      based on the the content above what does the song {song} by {artist} tell us about? give me a long explanations and
+      do not bold any text.
 
       """
-    prompt_template_song_meaning = PromptTemplate(input_variables=["artist", "song", "content"],
-                                                  template=template_song_meaning)
+    prompt_template_song_meaning = PromptTemplate(input_variables=["artist", "song", "content"], template=template_song_meaning)
     chain_song_meaning = LLMChain(llm=llm, prompt=prompt_template_song_meaning)
-    results_song_meaning = chain_song_meaning.run(artist=artist.title(), song=song.title(),
-                                                  content=docs[0].page_content)
+    results_song_meaning = chain_song_meaning.run(artist=artist.title(), song=song.title(), content=docs[0].page_content)
 
-    # get the song theme
-    template_song_theme = """
-      {artist} has released a song called {song}.
+    # get song recom
+    template_song_recom = """
+      here are the meaning of {song} by {artist}:
 
-      {content}
+      {song_meaning}
+
+      can you give me a 3 songs recommendation similar to the meaning of the song above?
+      use this format for the output and do not bold any text:
+      1. recommended song 1, with a brief explanation.
+      2. recommended song 2, with a brief explanation.
+      3. recommended song 3, with a brief explanation.
 
-      based on the the content above what themes does the lyrics have?
 
       """
-    prompt_template_song_theme = PromptTemplate(input_variables=["artist", "song", "content"],
-                                                template=template_song_theme)
-    chain_song_theme = LLMChain(llm=llm, prompt=prompt_template_song_theme)
-    text_song_theme = chain_song_theme.run(artist=artist.title(), song=song.title(), content=docs[0].page_content)
-    inputs_song_theme = tokenizer(text_song_theme, return_tensors="pt")
-    output_song_theme_proba = emotion_model(**inputs_song_theme).logits.softmax(1)
-    labels = emotion_model.config.id2label
-    confidences = {labels[i]: output_song_theme_proba[0][i].item() for i in range(len(labels))}
 
-    return results_song_meaning, confidences
+    prompt_template_song_recom = PromptTemplate(input_variables=["artist", "song", "song_meaning"], template=template_song_recom)
+    chain_song_recom = LLMChain(llm=llm, prompt=prompt_template_song_recom)
+    results_song_recom = chain_song_recom.run(artist=artist, song=song, song_meaning=results_song_meaning)
 
+    return results_song_meaning, results_song_recom
 
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     song = gr.Textbox(label="Song")
     artist = gr.Textbox(label="Artist")
     output_song_meaning = gr.Textbox(label="Meaning")
-    output_song_theme = gr.Label(num_top_classes=6, label="Theme")
-    gr.Interface(fn=song_insight, inputs=[song, artist], outputs=[output_song_meaning, output_song_theme])
-    example = gr.Examples([['Life Goes On', 'BTS'], ['Here Comes The Sun', 'The Beatles'],
-                           ['Bedtime Stories', 'Jay Chou'], ['Loser', 'BIGBANG']], [song, artist])
+    output_song_recom = gr.Textbox(label="Song Recommendation")
+    gr.Interface(fn=song_insight, inputs=[song, artist], outputs=[output_song_meaning, output_song_recom])
+    example = gr.Examples([["They Don't Care About Us", 'Michael Jackson'], ["Bad Romance", 'Lady Gaga'],
+                           ["Let It Be", "The Beatles"], ["Life Goes On", 'BTS'], ["Blank Space", "Taylor Swift"]], [song, artist])
 
 demo.launch()