Update app.py

app.py

@@ -24,7 +24,9 @@ import ast
 from utils import ClaudeLLM, ClaudeLLM2, extract_website_name, remove_numbers
 
 embeddings = HuggingFaceEmbeddings()
-db = FAISS.load_local('db_full', embeddings)
+
+db_art = FAISS.load_local('db_art', embeddings)
+db_yt = FAISS.load_local('db_yt', embeddings)
 
 mp_docs = {}
 llm_4 = ChatOpenAI(
@@ -32,6 +34,7 @@ llm_4 = ChatOpenAI(
             model='gpt-4'
         )
 claude = ClaudeLLM()
+claude2 = ClaudeLLM2()
 def add_text(history, text):
 
     print(history)
@@ -39,7 +42,12 @@ def add_text(history, text):
 
     return history, ""
 
-def retrieve_thoughts(query, ):
+def retrieve_thoughts(query, media):
+    if media[0] == "Articles":
+          db = db_art
+    else:
+      db = db_yt
+
   # print(db.similarity_search_with_score(query = query, k = k, fetch_k = k*10))
     docs_with_score = db.similarity_search_with_score(query = query, k = 1500, fetch_k = len(db.index_to_docstore_id.values()))
     df = pd.DataFrame([dict(doc[0])['metadata'] for doc in docs_with_score], )
@@ -48,19 +56,19 @@ def retrieve_thoughts(query, ):
 
   # TO-DO: What if user query doesn't match what we provide as documents
   # df.sort_values("score", inplace = True)
-    
+
     tier_1 = df
     tier_2 = df[((df['score'] < 1) * (df["score"] > 0.8))]
 
     tier_1
-    chunks_1 = tier_1.groupby(['title', 'url', ]).apply(lambda x: "\n...\n".join(x.sort_values('score').iloc[:3].sort_values('id')['page_content'].values)).values
+    chunks_1 = tier_1.groupby(['title', 'url', ]).apply(lambda x: "\n...\n".join(x.sort_values('id')['page_content'].values)).values
+    print(len(chunks_1[0]))
     score = tier_1.groupby(['title', 'url', ]).apply(lambda x: x.sort_values('score').iloc[:3]['score'].mean()).values
 
-    tier_1_adjusted = tier_1.groupby(['title', 'url', ]).first().reset_index()[[ 'title', 'url', 'author']]
+    tier_1_adjusted = tier_1.groupby(['title', 'url', ]).first().reset_index()[[ 'title', 'url']]
     tier_1_adjusted['content'] = chunks_1
     tier_1_adjusted['score'] = score
 
-
     chunks_2 = tier_2.groupby(['title', 'url', ]).apply(lambda x: "\n...\n".join(x.sort_values('id')['page_content'].values)).values
     tier_2_adjusted = tier_2.groupby(['title', 'url', ]).first().reset_index()[[ 'title', 'url']]
     tier_2_adjusted['content'] = chunks_2
@@ -70,88 +78,85 @@ def retrieve_thoughts(query, ):
     tier_1_adjusted.sort_values("score", inplace = True)
     tier_1_adjusted['ref'] = range(1, len(tier_1_adjusted) + 1 )
 
-    return {'tier 1':tier_1_adjusted[:min(len(tier_1_adjusted), 100)], 'tier 2': tier_2_adjusted.loc[:5]}
+    return {'tier 1':tier_1_adjusted[:min(len(tier_1_adjusted), 30)], 'tier 2': tier_2_adjusted.loc[:5]}
 
-def get_references(query):
+def get_references(query, media):
   # TO-DO FINSIH UPP.
-  thoughts = retrieve_thoughts(query)
+  thoughts = retrieve_thoughts(query, media)
   print(thoughts.keys())
   tier_1 = thoughts['tier 1']
-  reference = tier_1[['ref', 'url', 'title', 'author']].to_dict('records')
+  reference = tier_1[['ref', 'url', 'title']].to_dict('records')
   return reference
-def qa_themes(query,):
-
-    docs = ""
-
-    global db
-    print(db)
-
-    global mp_docs
-    thoughts = retrieve_thoughts(query)
-    if not(thoughts):
 
-        if mp_docs:
-            thoughts = mp_docs
-    else:
-        mp_docs = thoughts
-
-    tier_1 = thoughts['tier 1']
-    tier_2 = thoughts['tier 2']
+def grab_jsons(query, media = None, tier_1 = None, ):
 
-    reference = tier_1[['ref', 'url', 'title', 'author']].to_dict('records')
+  response = ""
 
-    tier_1 = list(tier_1.apply(lambda x: f"[{int(x['ref'])}] title: {x['title']}\n author: {x.author}\n content: {x.content}", axis = 1).values)
-    tier_2 = list(tier_2.apply(lambda x: f"title: {x['title']}\n Content: {x.content}", axis = 1).values)
-    print(f"QUERY: {query}\nTIER 1: {tier_1}\nTIER2: {tier_2}")
+  if tier_1 is None:
+      thoughts = retrieve_thoughts(query, media)
+      tier_1 = thoughts['tier 1']
+      tier_1 = list(tier_1.apply(lambda x: f"[{int(x['ref'])}] title: {x['title']}\n Content: {x.content}", axis = 1).values)
 
-    # print(f"DOCS RETRIEVED: {mp_docs.values}")
-    # Cynthesis Generation
+  for i in range(3, len(tier_1), 3):
+    portion = tier_1[i - 3 :i]
+    response += '\n' + jsonify_articles(query, portion)
 
+  return response
 
-    # Themes
-    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will use do different analysis to the articles provided to me. Stay truthful and if you weren't provided any resources give your oppinion only."""
-    task = """Your primary responsibility is to identify multiple themes from the given articles. For each theme detected, you are to present it under three separate categories:
+def jsonify_articles(query, tier_1 = None):
+    if tier_1 is None:
+      thoughts = retrieve_thoughts(query)
+      tier_1 = thoughts['tier 1']
+      tier_1 = list(tier_1.apply(lambda x: f"[{int(x['ref'])}] title: {x['title']}\n Content: {x.content}", axis = 1).values)
 
-    1. Theme Title - An easy-to-understand title that encapsulates the core idea of the theme extracted from the article.
-
-    2. Theme Description - An expanded elaboration that explores the theme in detail based on the arguments and points provided in the article.
-
-    3. Quotes related to theme - Locate and provide at least one compelling quote from the article that directly supports or showcases the theme you have identified. This quote should serve as a specific evidence or example from the article text that corresponds directly to the developed theme.
-    
+        # json
+#         {
+#     'ref': 1,
+#     'quotes': ['quote_1', 'quote_2', 'quote_3'],
+#     'summary (optional for now as we already have summaries)': ""
+# }
 
-    Keep your answer direct and don't include your thoughts. Make sure that the quote used should have a reference [1] related to the article."""
+    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will execute different analysis to the articles provided to you. Stay truthful and if you weren't provided any resources give your oppinion only."""
+    task = """Your primary responsibility is to identify valuable information from the given articles related to a given query.
+    For each article provided, you are to present it under four separate categories:
+    1. Article Reference - A reference for the article id: int
+    2. Article Title - The title for the article: string
+    3. Article quotes - Numerous Quotes extracted from the article that prove certain point of views in a list format [quote_1, quote_2, quote_3, quote_4, quote_5]
+    4. Article Summary - A summary for the article: string
 
+    Make sure to include all valuable quotes to be used later on.
+    Keep your answer direct and don't include your thoughts. Make sure that the quote used should have a reference [1] that identifies the source."""
 
     prompt = PromptTemplate(
-            input_variables=["query", "task", "session_prompt", "articles"],
+            input_variables=["query", "task", "articles"],
             template="""
-            You are a {session_prompt}
             {task}
+
+            The extracted information should correlate to the following query.
             query: {query}
             Articles:
             {articles}
 
-            The extracted themes should be written in structured manner, ensuring clarity and meaningful correlation between the themes and the articles. Don't forget to mention the reference in the quote. Avoid including personal opinions or making generalizations that are not explicitly supported by the articles.
+            The extracted information should be written in structured manner, ensuring clarity and meaningful format for the articles. Avoid including personal opinions or making generalizations that are not explicitly supported by the articles.
             Keep your answer direct and don't include your thoughts.
             """,
         )
 
-
-    # llm = BardLLM()
     chain = LLMChain(llm=claude, prompt = prompt)
 
-    themes = chain.run(query=query, articles="\n".join(tier_1), session_prompt = session_prompt, task = task)
-    return themes
+    json_articles = chain.run(query=query, articles="\n".join(tier_1), task = task).strip()
 
-def qa_retrieve(query,):
+    return json_articles
 
+def qa_retrieve(query, media):
+    
     docs = ""
 
     global db
     print(db)
 
     global mp_docs
-    thoughts = retrieve_thoughts(query)
+    thoughts = retrieve_thoughts(query, media)
     if not(thoughts):
 
         if mp_docs:
@@ -164,14 +169,14 @@ def qa_retrieve(query,):
 
     reference = tier_1[['ref', 'url', 'title']].to_dict('records')
 
-    tier_1 = list(tier_1.apply(lambda x: f"[{int(x['ref'])}] title: {x['title']}\n author: {x.author}\n content: {x.content}", axis = 1).values)
+    tier_1 = list(tier_1.apply(lambda x: f"ref: [{int(x['ref'])}]\ntitle: {x['title']}\n Content: {x.content}", axis = 1).values)
     tier_2 = list(tier_2.apply(lambda x: f"title: {x['title']}\n Content: {x.content}", axis = 1).values)
     print(f"QUERY: {query}\nTIER 1: {tier_1}\nTIER2: {tier_2}")
 
     # print(f"DOCS RETRIEVED: {mp_docs.values}")
     # Cynthesis Generation
 
-    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will use do different analysis to the articles provided to me. Stay truthful and if you weren't provided any resources give your oppinion only."""
+    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will use do different analysis to the articles provided to you. You will generate a rich synthesis with numerous arguments extracted from the sources given."""
     # task = """Create a coherent synthesis in which you use references to the id of articles provided and relevant to the query.
 
     # Follow the example structure:
@@ -184,41 +189,42 @@ def qa_retrieve(query,):
     # - Popular wine choices include Cabernet Sauvignon, Pinot Noir, Zinfandel, Malbec, Syrah, and Merlot [2].
     # Remember, the goal is to choose a wine that complements the cut of steak and not overwhelm or take away from the flavor of the meat [3]."
     # """
-
+    articles = grab_jsons(query, tier_1 = tier_1)
     prompt = PromptTemplate(
             input_variables=["query", "session_prompt", "articles"],
             template="""
           You are a {session_prompt}
-          Create a coherent well-structured synthesis in which you use references to the id of articles provided and relevant to the query.
+          Create a rich well-structured synthesis in which you use references to the id of articles provided and relevant to the query.
+          The Synthesis should include at least 500 words englobing all prespectives found in the sources.
 
-              Follow the example structure, references are not provided but are found in the answer:
-              User: What are the secondary effects of covid?
-              Cynthesis: \nSecondary effects of COVID-19, often referred to as \"Long COVID\", are a significant concern. These effects are not limited to the acute phase of the disease but persist well past the first month, affecting various organ systems and leading to adverse outcomes such as all-cause death and hospitalization [1]. \n\nOne of the most alarming secondary effects is the increased risk of cardiovascular diseases. Studies have shown a 1.6-fold increased risk of stroke and a 2-fold higher risk of acute coronary disease in individuals who had COVID-19 [2][3][8]. These risks were observed even in younger populations, with a mean age of 44, and were prevalent after 30 days post-infection [2][3]. \n\nAnother study found that the adverse outcomes of COVID-19 could persist up to the 2-year mark, with the toll of adverse sequelae being worst during the first year [3]. The study also highlighted that individuals with severe COVID-19, who were hospitalized, were more likely to be afflicted with protracted symptoms and new medical diagnoses [3]. \n\nHowever, it's important to note that the risks associated with Long COVID might be most significant in the first few weeks post-infection and fade away as time goes on [4][9]. For instance, the chance of developing pulmonary embolism was found to be 32 times higher in the first month after testing positive for COVID-19 [4]. \n\nMoreover, the number of excess deaths in the U.S., which would indicate fatal consequences of mild infections at a delay of months or years, dropped to zero in April, about two months after the end of the winter surge, and have stayed relatively low ever since [4]. This suggests that a second wave of deaths—a long-COVID wave—never seems to break [4]. \n\nIn conclusion, while the secondary effects of COVID-19 are significant and can persist for a long time, the most severe risks seem to occur in the first few weeks post-infection and then gradually decrease. However, the full extent of the long-term effects of COVID-19 is still unknown, and further research is needed to fully understand the ways and extent COVID-19 has affected us.",
+          Follow the example structure:
+          User: What are the secondary effects of covid?
+          Synthesis: \nSecondary effects of COVID-19, often referred to as \"Long COVID\", are a significant concern. These effects are not limited to the acute phase of the disease but persist well past the first month, affecting various organ systems and leading to adverse outcomes such as all-cause death and hospitalization [1]. \n\nOne of the most alarming secondary effects is the increased risk of cardiovascular diseases. Studies have shown a 1.6-fold increased risk of stroke and a 2-fold higher risk of acute coronary disease in individuals who had COVID-19 [2][3][8]. These risks were observed even in younger populations, with a mean age of 44, and were prevalent after 30 days post-infection [2][3]. \n\nAnother study found that the adverse outcomes of COVID-19 could persist up to the 2-year mark, with the toll of adverse sequelae being worst during the first year [3]. The study also highlighted that individuals with severe COVID-19, who were hospitalized, were more likely to be afflicted with protracted symptoms and new medical diagnoses [3]. \n\nHowever, it's important to note that the risks associated with Long COVID might be most significant in the first few weeks post-infection and fade away as time goes on [4][9]. For instance, the chance of developing pulmonary embolism was found to be 32 times higher in the first month after testing positive for COVID-19 [4]. \n\nMoreover, the number of excess deaths in the U.S., which would indicate fatal consequences of mild infections at a delay of months or years, dropped to zero in April, about two months after the end of the winter surge, and have stayed relatively low ever since [4]. This suggests that a second wave of deaths—a long-COVID wave—never seems to break [4]. \n\nIn conclusion, while the secondary effects of COVID-19 are significant and can persist for a long time, the most severe risks seem to occur in the first few weeks post-infection and then gradually decrease. However, the full extent of the long-term effects of COVID-19 is still unknown, and further research is needed to fully understand the ways and extent COVID-19 has affected us.",
 
             query: {query}
 
             Articles:
             {articles}
 
-            Make sure to quote the article used if the argument corresponds to the query.
-            Use careful reasoning and professional writing for the synthesis. No need to mention your interaction with articles.
+            Make sure to quote the article used if the argument corresponds to the query. Add as many arguments and point of views as possible as long as they have a reference.
+            Keep your answer direct and use careful reasoning and professional writing for the synthesis. No need to mention your interaction with articles.
             Remember not to mention articles used at the beginning of sentences, keep it cohesive and rich in text while referencing as much as possible of sources given.
             """,
         )
 
 
-    llm = ClaudeLLM2()
-    chain = LLMChain(llm=llm, prompt = prompt)
+    # llm = BardLLM()
+    chain = LLMChain(llm=claude2, prompt = prompt)
+
+    consensus = chain.run(query=query, articles=articles, session_prompt = session_prompt,)
+    consensus = consensus[consensus.index(':')+1:].strip()
 
-    consensus = chain.run(query=query, articles="\n".join(tier_1), session_prompt = session_prompt,).strip()
     if "In conclusion" in consensus:
       consensus = consensus[:consensus.index('In conclusion')]
-        
-    consensus = consensus[consensus.index(':')+1:].strip()
-    
-    intro = qa_intro(query, consensus, tier_1)
-    conclusion = qa_conclusion(query, consensus, tier_1)
-    cynthesis = intro + "\n\n" + consensus + conclusion
+    intro = qa_intro(query, consensus, articles)
+    conclusion = qa_conclusion(query, consensus, articles)
+    cynthesis = intro + "\n\n" + consensus + "\n\n" + conclusion
+
     # for i in range(1, len(tier_1)+1):
     #   response = response.replace(f'[{i}]', f"<span class='text-primary'>[{i}]</span>")
 
@@ -226,41 +232,40 @@ def qa_retrieve(query,):
     # json_resp = {'cynthesis': response, 'questions': questions, 'reference': reference}
 
     return cynthesis
+
 def qa_intro(query, cynthesis, tier_1,):
-  
+
 
   llm = ClaudeLLM()
   llm_4 = ChatOpenAI(
-              temperature=0.5,
+              temperature=0,
               model='gpt-3.5-turbo-16k'
           )
 
   session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will use do different analysis to the articles provided to me. Stay truthful and if you weren't provided any resources give your oppinion only."""
 
   prompt = PromptTemplate(
-          input_variables=["query", "articles"],
+          input_variables=["query", "cynthesis", "articles"],
           template="""
-          Give me an introduction to the following topic. Consider this an abstract. And after finishing the introduction, pick one quote from the sources given below.
-          this is the desired structure:
-          Introduction: Introduction
-          Quote: "quote". [1] (Reference)
+          Give me an introduction to the following consensus without repeating how it starts. Consider this an abstract. And after finishing the introduction, pick one quote from the sources given below.
 
-          Follow the structure given.
           query: {query}
-          
+          Here's the consensus: {cynthesis}
+
           We have the opportunity to give an introduction to this synthesis without repeating information found.
           Pick an opening quote from the sources given below\n
           ---------\n
           {articles}
           ---------\n
 
-          Don't forget that your job is to only provide an introduction, abstract part that introduces the synthesis and then pick one general quote from the sources given. Maintain the desired structure.""",
+          Don't forget that your job is to only provide an introduction, abstract part that introduces the synthesis without repeating it and then pick one general quote from the sources given.""",
       )
 
+
   # llm = BardLLM()
   chain = LLMChain(llm=llm_4, prompt = prompt)
 
-  intro = chain.run(query=query, articles="\n".join(tier_1[:15]))
+  intro = chain.run(query=query, articles=tier_1, cynthesis = cynthesis)
   return intro.strip()
 
 def qa_conclusion(query, cynthesis, tier_1,):
@@ -295,12 +300,12 @@ def qa_conclusion(query, cynthesis, tier_1,):
   # llm = BardLLM()
   chain = LLMChain(llm=llm_4, prompt = prompt)
 
-  conclusion = chain.run(query=query, articles="\n".join(tier_1[:15]), cynthesis = cynthesis)
+  conclusion = chain.run(query=query, articles=tier_1, cynthesis = cynthesis)
   return conclusion.strip()
-    
-def qa_faqs(query):
 
-    thoughts = retrieve_thoughts(query)
+def qa_faqs(query, media):
+
+    thoughts = retrieve_thoughts(query, media)
 
     # tier_1 = thoughts['tier 1']
     tier_2 = thoughts['tier 2']
@@ -310,7 +315,7 @@ def qa_faqs(query):
     # tier_1 = list(tier_1.apply(lambda x: f"[{int(x['ref'])}] title: {x['title']}\n Content: {x.content}", axis = 1).values)
     tier_2 = list(tier_2.apply(lambda x: f"title: {x['title']}\n Content: {x.content}", axis = 1).values)
 # Generate related questions
-    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will use do different analysis to the articles provided to me. Stay truthful and if you weren't provided any resources give your oppinion only."""
+    session_prompt = """ A bot that is open to discussions about different cultural, philosophical and political exchanges. You will do different analysis to the articles provided to me. Stay truthful and if you weren't provided any resources give your oppinion only."""
 
     prompt_q = PromptTemplate(
             input_variables=[ "session_prompt", "articles"],
@@ -331,16 +336,16 @@ def qa_faqs(query):
     questions = questions[questions.index('1'):]
 
     questions = [ t.strip() for (i, t) in enumerate(questions.split('\n\n'))  if len(t) > 5][:5]
-    
+
     return "\n\n".join(questions)
-# examples = [
-#     ["Will Russia win the war in Ukraine?"],
+examples = [
+    ["Will Russia win the war in Ukraine?"],
 
-#     ]
+    ]
 
-# demo = gr.Interface(fn=qa_retrieve, title="cicero-qa-api",
-#                      inputs=gr.inputs.Textbox(lines=5, label="what would you like to learn about?"),
-#                      outputs="json",examples=examples)
+demo = gr.Interface(fn=qa_retrieve, title="cicero-qa-api",
+                     inputs=gr.inputs.Textbox(lines=5, label="what would you like to learn about?"),
+                     outputs="json",examples=examples)
 
 def parallel_greet_claude(batch, ):
 
@@ -356,19 +361,19 @@ def parallel_greet_claude(batch, ):
 #     ["Covid Global Impact and what's beyond that"]
 #     ]
 
-cynthesis = gr.Interface(fn = qa_retrieve, inputs = "text", outputs = gr.components.Textbox(lines=3, label="Cynthesis"))
-questions = gr.Interface(fn = qa_faqs, inputs = "text", outputs = gr.components.Textbox(lines=3, label="Related Questions"))
-themes = gr.Interface(fn = qa_themes, inputs = "text", outputs = gr.components.Textbox(lines=3, label="themes"))
+cynthesis = gr.Interface(fn = qa_retrieve, inputs = ["text", gr.CheckboxGroup(["Articles", "Podcasts", "Youtube"], label="Media", info="Choose One Type of Media until we merge (Podcasts excluded for now)"),], outputs = gr.components.Textbox(lines=3, label="Cynthesis"))
+questions = gr.Interface(fn = qa_faqs, inputs = ["text", gr.CheckboxGroup(["Articles", "Podcasts", "Youtube"], label="Media", info="Choose One Type of Media until we merge (Podcasts excluded for now)"),], outputs = gr.components.Textbox(lines=3, label="Related Questions"))
+#themes = gr.Interface(fn = qa_themes, inputs = ["text", gr.CheckboxGroup(["Articles", "Podcasts", "Youtube"], label="Media", info="Choose One Type of Media until we merge (Podcasts excluded for now)"),], outputs = gr.components.Textbox(lines=3, label="themes"))
 
 # gpt_3 = gr.Interface(fn = parallel_greet_gpt_3, inputs = "text", outputs = gr.components.Textbox(lines=3, label="GPT3.5"))
 # gpt_4 = gr.Interface(fn = parallel_greet_gpt_4, inputs = "text", outputs = gr.components.Textbox(lines=3, label="GPT4"))
 # claude = gr.Interface(fn = parallel_greet_claude, inputs = "text", outputs = gr.components.Textbox(lines=3, label="Claude"))
-reference = gr.Interface(fn = get_references, inputs = "text", outputs = "json", label = "Reference")
-
-demo = gr.Parallel(cynthesis, themes, questions, reference)
-
-# demo = gr.Series(references, cynthesis, themes)
+reference = gr.Interface(fn = get_references, inputs = ["text", gr.CheckboxGroup(["Articles", "Podcasts", "Youtube"], label="Media", info="Choose One Type of Media until we merge (Podcasts excluded for now)"),], outputs = "json", label = "Reference")
+json = gr.Interface(fn = grab_jsons, inputs = ["text", gr.CheckboxGroup(["Articles", "Podcasts", "Youtube"], label="Media", info="Choose One Type of Media until we merge (Podcasts excluded for now)"),], outputs = gr.components.Textbox(lines=3, label="json"))
+#demo = gr.Parallel(cynthesis, reference)
+#demo = gr.Parallel(themes, reference)
 
+demo = gr.Parallel(json, cynthesis, questions, reference)
 demo.queue(concurrency_count = 4)
 demo.launch()