Upload 5 files

app.py

@@ -7,7 +7,7 @@ import gradio as gr
 import main
 
 # Define two separate functions for each button
-def call_generate_news(subject, sites, min_words, max_words):
+def call_generate_news(subject, sites, min_words, max_words, model):
     if subject == '':
         return 'Erro: Assunto não informado', ''
     if min_words == '':
@@ -24,11 +24,11 @@ def call_generate_news(subject, sites, min_words, max_words):
     for item in list_sites:
         if item:
             filtered_list_sites.append(item)
-    result_news = main.generate_news(subject, min_words, max_words, filtered_list_sites)
+    result_news = main.generate_news(subject, min_words, max_words, filtered_list_sites, model.lower())
     return result_news
 
-def call_invoke_llm(context, prompt):
-    result = main.call_llm(context, prompt)
+def call_invoke_llm(context, prompt, model):
+    result = main.call_llm(context, prompt, model)
     return result
 
 # Create the Gradio interface using Blocks
@@ -41,11 +41,16 @@ with gr.Blocks(title='BotNews') as page:
     with gr.Row():
         input1 = gr.Textbox(label="Assunto:", lines=1)
     with gr.Row():
+        input5 = gr.Dropdown(
+            ["OpenAI", "Mixtral-8x7B"], value='OpenAI', label="Modelo LLM", info="Escolha o modelo para processamento das informações.",
+             scale=1
+        )
         input2 = gr.Textbox(label="Domínios para pesquisa (coloque um domínio por linha):",
-                            lines=10, value='https://www.cnnbrasil.com.br/\nhttps://g1.globo.com/\nhttps://www.metropoles.com/\nhttps://www.bbc.com/portuguese/\nhttps://www.instagram.com/')
-    with gr.Row():
-        input3 = gr.Textbox(label="Mínimo de palavras:", lines=1, value="300")
-        input4 = gr.Textbox(label="Máximo de palavras:", lines=1, value="700")
+                            lines=8,
+                            scale=2,
+                            value='https://www.cnnbrasil.com.br/\nhttps://g1.globo.com/\nhttps://www.metropoles.com/\nhttps://www.bbc.com/portuguese/\nhttps://www.instagram.com/')
+        input3 = gr.Textbox(label="Mínimo de palavras:", lines=1, value="300", scale=1)
+        input4 = gr.Textbox(label="Máximo de palavras:", lines=1, value="700", scale=1)
     with gr.Row():
         button1 = gr.Button("Gerar notícia")
     with gr.Row():
@@ -56,14 +61,14 @@ with gr.Blocks(title='BotNews') as page:
     gr.Markdown("<b>Instrução:</b> Preencha abaixo com um comando para ser executado sobre a notícia (Por exemplo: 'Resuma em tópicos' ou 'Adicione um tom sarcástico').")
     gr.Markdown(" ")
     with gr.Row():
-        input5 = gr.Textbox(label="Prompt para a notícia")
+        input6 = gr.Textbox(label="Prompt para a notícia")
     with gr.Row():
         button2 = gr.Button("Gerar resposta")
     with gr.Row():
         output2 = gr.Textbox(label="Resposta gerada por IA", lines=25)
 
-    button1.click(call_generate_news, inputs=[input1, input2, input3, input4], outputs=[output1])
-    button2.click(call_invoke_llm, inputs=[output1, input5], outputs=[output2])
+    button1.click(call_generate_news, inputs=[input1, input2, input3, input4, input5], outputs=[output1])
+    button2.click(call_invoke_llm, inputs=[output1, input6, input5], outputs=[output2])
 
 # Launch the interface
 page.launch() #share=True

llm.py

@@ -2,9 +2,11 @@
 # Description: 
 # ------------------------------------------------
 from langchain_openai import ChatOpenAI
+from langchain_together import Together
+
 import os
 
-def invoke_llm(context, task):
+def invoke_llm(context, task, model):
     prompt = f"""You are an assistant of a newspaper. 
                  Execute the task just based on the given context.
                  The task is delimited by <> and the context is delimited by <>.
@@ -15,10 +17,17 @@ def invoke_llm(context, task):
                  Answer here:
               """
 
-    llm=ChatOpenAI(model_name="gpt-3.5-turbo",
-                       temperature=0.3,
-                       openai_api_key=os.environ['OPENAI_KEY'],
-                       max_tokens=1000)
+    if model == 'openai':
+        llm=ChatOpenAI(model_name="gpt-3.5-turbo",
+                           temperature=0.3,
+                           openai_api_key=os.environ['OPENAI_KEY'],
+                           max_tokens=1000)
+    else:
+        llm=ChatOpenAI(model_name="mistralai/Mixtral-8x7B-Instruct-v0.1",
+                           temperature=0.3,
+                           together_api_key=os.environ['TOGETHER_KEY'],
+                           max_tokens=1000)
+
     result = llm.invoke(prompt)
     return result.content
 

main.py

@@ -8,7 +8,7 @@ import constants
 import llm
 
 
-def generate_news(subject, min_words, max_words, sites):
+def generate_news(subject, min_words, max_words, sites, model):
     print('\n\n' + '*' * 50)
     print('\n\nInício do Programa: \n')
 
@@ -31,13 +31,13 @@ def generate_news(subject, min_words, max_words, sites):
         return 'Erro: ' + ret
 
     print('\nGerando embeddings e vectorstore...')
-    vectorstore = rag.generate_embeddings_and_vectorstore(constants.local_base)
+    vectorstore = rag.generate_embeddings_and_vectorstore(constants.local_base, model)
     if type(vectorstore) == str:
         return 'Erro: ' + vectorstore
 
     print('\nGerando a notícia (RAG)...')
     print('    Assunto: ' + subject)
-    obj_rag = rag.Rag(vectorstore, min_words, max_words)
+    obj_rag = rag.Rag(vectorstore, min_words, max_words, model)
     result_news = obj_rag.generate_text(subject)
     if type(result_news) == str:
         if 'maximum context length' in result_news:
@@ -56,8 +56,8 @@ def generate_news(subject, min_words, max_words, sites):
     else:
         return result_news[0]
 
-def call_llm(context, prompt):
+def call_llm(context, prompt, model):
     print('\nChamando o modelo de linguagem...')
-    result = llm.invoke_llm(context, prompt)
+    result = llm.invoke_llm(context, prompt, model)
 
     return result

rag.py

@@ -3,11 +3,13 @@
 # ------------------------------------------------
 #from langchain.embeddings import OpenAIEmbeddings
 from langchain_openai import OpenAIEmbeddings
+from langchain_community.embeddings import HuggingFaceEmbeddings
 from langchain_community.vectorstores import Chroma
 from langchain_community.document_loaders import DirectoryLoader
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain.prompts import PromptTemplate
 from langchain_openai import ChatOpenAI
+from langchain_together import Together
 from langchain.memory import ConversationBufferMemory
 from langchain.chains import ConversationalRetrievalChain
 import os
@@ -22,7 +24,7 @@ def read_csv_to_dict(filename):
             data_dict[key] = value
     return data_dict
 
-def generate_embeddings_and_vectorstore(path):
+def generate_embeddings_and_vectorstore(path, model):
     try:
         loader = DirectoryLoader(path=path, glob="**/*.txt")
         corpus = loader.load()
@@ -41,8 +43,10 @@ def generate_embeddings_and_vectorstore(path):
 
         #print('docs')
         #print(docs)
-
-        fc_embeddings = OpenAIEmbeddings(openai_api_key=os.environ['OPENAI_KEY'])
+        if model == 'openai':
+            fc_embeddings = OpenAIEmbeddings(openai_api_key=os.environ['OPENAI_KEY'])
+        else:
+            fc_embeddings = HuggingFaceEmbeddings(model_name = 'intfloat/multilingual-e5-large-instruct')
         vectorstore = Chroma.from_documents(docs, fc_embeddings)
         print('total de docs no vectorstore=',len(vectorstore.get()['documents']))
 
@@ -52,7 +56,7 @@ def generate_embeddings_and_vectorstore(path):
         return str(e)
 
 class Rag:
-    def __init__(self, vectorstore, min_words, max_words):
+    def __init__(self, vectorstore, min_words, max_words, model):
         self.text = None
         self.vectorstore = vectorstore
         self.memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, output_key="answer")
@@ -72,20 +76,35 @@ class Rag:
                                      input_variables=["context", "question"],
                                      partial_variables={"min_words": min_words, "max_words": max_words})
 
-        self.qa = ConversationalRetrievalChain.from_llm(
-                    llm=ChatOpenAI(model_name="gpt-3.5-turbo-0125", #0125 #1106
-                                   temperature=0,
-                                   openai_api_key=os.environ['OPENAI_KEY'],
-                                   max_tokens=int(int(max_words) + (int(max_words) / 2))), #número máximo de tokens para a resposta
-                    memory=self.memory,
-                    # retriever=vectorstore.as_retriever(search_type='similarity_score_threshold',
-                    #                                    search_kwargs={'k':4, 'score_threshold':0.8}), #search_kwargs={'k': 3}
-                    retriever=vectorstore.as_retriever(),
-                    combine_docs_chain_kwargs={"prompt": self.prompt},
-                    chain_type="stuff",#map_reduce, refine, map_rerank
-                    return_source_documents=True,
-                )
-        
+        if model == 'openai':
+            self.qa = ConversationalRetrievalChain.from_llm(
+                        llm=ChatOpenAI(model_name="gpt-3.5-turbo-0125", #0125 #1106
+                                       temperature=0,
+                                       openai_api_key=os.environ['OPENAI_KEY'],
+                                       max_tokens=int(int(max_words) + (int(max_words) / 2))), #número máximo de tokens para a resposta
+                        memory=self.memory,
+                        # retriever=vectorstore.as_retriever(search_type='similarity_score_threshold',
+                        #                                    search_kwargs={'k':4, 'score_threshold':0.8}), #search_kwargs={'k': 3}
+                        retriever=vectorstore.as_retriever(),
+                        combine_docs_chain_kwargs={"prompt": self.prompt},
+                        chain_type="stuff",#map_reduce, refine, map_rerank
+                        return_source_documents=True,
+                    )
+        else:
+            self.qa = ConversationalRetrievalChain.from_llm(
+                        llm=Together(model="mistralai/Mixtral-8x7B-Instruct-v0.1", #0125 #1106
+                                       temperature=0,
+                                       together_api_key=os.environ['TOGETHER_KEY'],
+                                       max_tokens=int(int(max_words) + (int(max_words) / 2))), #número máximo de tokens para a resposta
+                        memory=self.memory,
+                        # retriever=vectorstore.as_retriever(search_type='similarity_score_threshold',
+                        #                                    search_kwargs={'k':4, 'score_threshold':0.8}), #search_kwargs={'k': 3}
+                        retriever=vectorstore.as_retriever(),
+                        combine_docs_chain_kwargs={"prompt": self.prompt},
+                        chain_type="stuff",#map_reduce, refine, map_rerank
+                        return_source_documents=True,
+                    )
+
         # from langchain_together import Together
         # self.qa = ConversationalRetrievalChain.from_llm(
         #     llm=Together(model="mistralai/Mixtral-8x7B-Instruct-v0.1",  # 0125 #1106

requirements.txt

@@ -2,8 +2,10 @@ google-api-python-client
 langchain
 langchain-community
 langchain_openai
+langchain-together
 openai
 unstructured
 chromadb
 tiktoken
-gradio
+gradio
+sentence-transformers