yiha

app.py

@@ -23,8 +23,6 @@ app.add_middleware(
 )
 
 
-# Define retriever_chain as a global variable
-retriever_chain = None
 
 # Function to crawl all URLs from a domain
 def get_all_links_from_domain(domain_url):
@@ -47,9 +45,9 @@ def get_links_from_page(url, visited_urls, all_links, base_domain):
         return
 
     visited_urls.add(url)
-    print("Getting next" + url)
+    print("Getting next " + url)
     response = requests.get(url)
-
+    print(response)
     if response.status_code == 200:
         soup = BeautifulSoup(response.content, 'html.parser')
         base_url = urlparse(url).scheme + '://' + urlparse(url).netloc
@@ -59,29 +57,39 @@ def get_links_from_page(url, visited_urls, all_links, base_domain):
             href = link.get('href')
             absolute_url = urljoin(base_url, href)
             parsed_url = urlparse(absolute_url)
-            
-            if parsed_url.netloc == base_domain:
-                all_links.add(absolute_url)
-                get_links_from_page(absolute_url, visited_urls, all_links, base_domain)
+            if absolute_url not in visited_urls: 
+              if absolute_url.startswith(base_domain):
+                  print("hrefe " +absolute_url)
+                  all_links.add(absolute_url)
+                  get_links_from_page(absolute_url, visited_urls, all_links, base_domain)
 
     else:
         print(f"Failed to retrieve content from {url}. Status code: {response.status_code}")
 
-# Function to load the RAG model
-def load_rag_model():
-    model =  HuggingFaceHub(
-        repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
-        model_kwargs={"max_length": 1048, "temperature":0.2, "max_new_tokens":256, "top_p":0.95, "repetition_penalty":1.0},
-    )
-    return model
+
+from langchain_community.embeddings import HuggingFaceBgeEmbeddings 
+from langchain.chains import RetrievalQA
+
+from langchain.memory import ConversationBufferMemory
+
+
 
 # Function to index URLs in RAG
-def index_urls_in_rag(urls):
+def index_urls_in_rag(urls=[]):
+      # Load the RAG model
+    rag_model = "BAAI/bge-base-en-v1.5"
+    encode_kwargs = {
+        "normalize_embeddings": True
+    }  # set True to compute cosine similarity
+    embeddings = HuggingFaceBgeEmbeddings(
+        model_name=rag_model, encode_kwargs=encode_kwargs, model_kwargs={"device": "cpu"}
+    )
+
     # Create a vector store for storing embeddings of documents
-    vector_store = Chroma(persist_directory="/home/user/.cache/chroma_db")
+    vector_store = Chroma(persist_directory="/home/user/.cache/chroma_db", embedding_function=embeddings)
     
-    # Load the RAG model
-    rag_model = load_rag_model()
+    print(urls)
+
 
     for url in urls:
         # Get text from the URL
@@ -95,37 +103,56 @@ def index_urls_in_rag(urls):
         # Index document chunks into the vector store
         vector_store.add_documents(document_chunks)
 
+    
     # Convert vector store to retriever
     retriever = vector_store.as_retriever()
+ 
+    return retriever
+
+
+
+# Function to load the RAG model
+def load_model():
+    model =  HuggingFaceHub(
+        repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
+        model_kwargs={"max_length": 1048, "temperature":0.1, "max_new_tokens":512, "top_p":0.95, "repetition_penalty":1.0},
+    )
+    return model
 
-    # Define prompt for RAG model
+def get_conversational_rag_chain(retriever_chain): 
+    
+    llm = load_model()
+    
     prompt = ChatPromptTemplate.from_messages([
+      ("system", "Du bist eine freundlicher Mitarbeiterin Namens Susie und arbeitest in einenm Call Center. Antworte mit: 'Ich bin mir nicht sicher. Wollen Sie eine Mitarbeiter sprechen' Wenn die Antwort nicht aus dem Context hervorgeht. Antworte bitte immer auf Deutsch? CONTEXT:\n\n{context}"),
       MessagesPlaceholder(variable_name="chat_history"),
       ("user", "{input}"),
     ])
-
-    # Create history-aware retriever chain
-    retriever_chain = create_history_aware_retriever(rag_model, retriever, prompt)
-
-    return retriever_chain
-
-# Index URLs on app startup
-@app.on_event("startup")
-async def startup():
-    domain_url = 'https://www.bofrost.de/faq/'
-    urls = get_all_links_from_domain(domain_url)
     
-    print(urls)
+    stuff_documents_chain = create_stuff_documents_chain(llm,prompt)
+    
+    return create_retrieval_chain(retriever_chain, stuff_documents_chain)
 
-# Define API endpoint to receive queries and provide responses
-@app.post("/generate/")
-def generate(user_input):
+    
+def get_response(message, history=[]): 
  
-    return get_response(user_input, [])
+    conversation_rag_chain = get_conversational_rag_chain(retriever_chain)
+    
+    response = conversation_rag_chain.invoke({
+        "chat_history": history,
+        "input": message  + " Assistant: ",
+        "chat_message": message + " Assistant: "
+    })
+    #print("get_response " +response)
+    res = response['answer']
+    parts = res.split(" Assistant: ")
+    last_part = parts[-1]
+    return last_part#[-1]['generation']['content']   
 
+ 
 def get_conversational_rag_chain(retriever_chain): 
     
-    llm = load_model(model_name)
+    llm = load_model()
     
     prompt = ChatPromptTemplate.from_messages([
       ("system", "Du bist eine freundlicher Mitarbeiterin Namens Susie und arbeitest in einenm Call Center. Du beantwortest basierend auf dem Context. Benutze nur den Inhalt des Context. Füge wenn möglich die Quelle hinzu. Antworte mit: Ich bin mir nicht sicher. Wenn die Antwort nicht aus dem Context hervorgeht. Antworte auf Deutsch, bitte? CONTEXT:\n\n{context}"),
@@ -138,30 +165,8 @@ def get_conversational_rag_chain(retriever_chain):
     return create_retrieval_chain(retriever_chain, stuff_documents_chain)
 
     
-def get_response(message, history=[]):
-   # dialog = history_to_dialog_format(history)
-   # dialog.append({"role": "user", "content": message})
-
-      # Define the prompt as a ChatPromptValue object
-    #user_input = ChatPromptValue(user_input)
-    
-    # Convert the prompt to a tensor
-    #input_ids = user_input.tensor
-    
-
-    model = "BAAI/bge-base-en-v1.5"
-    encode_kwargs = {
-        "normalize_embeddings": True
-    }  # set True to compute cosine similarity
-    embeddings = HuggingFaceBgeEmbeddings(
-        model_name=model, encode_kwargs=encode_kwargs, model_kwargs={"device": "cpu"}
-    )
-    # load from disk
-    vector_store = Chroma(persist_directory="/home/user/.cache/chroma_db", embedding_function=embeddings)
+def get_response(message, history=[]): 
  
-  
-    history =[]
-    retriever_chain = get_context_retriever_chain(vector_store)
     conversation_rag_chain = get_conversational_rag_chain(retriever_chain)
     
     response = conversation_rag_chain.invoke({
@@ -174,3 +179,18 @@ def get_response(message, history=[]):
     parts = res.split(" Assistant: ")
     last_part = parts[-1]
     return last_part#[-1]['generation']['content']   
+
+
+# Index URLs on app startup
+@app.on_event("startup")
+async def startup():
+    domain_url = 'https://www.bofrost.de/faq/'
+    links = get_all_links_from_domain(domain_url)
+    retriever_chain = index_urls_in_rag(links)
+    retriever_chain.invoke("Was ist bofrost*")
+    get_response("Wlautet die URL Schoko Osterhasen?")
+
+# Define API endpoint to receive queries and provide responses
+@app.post("/generate/")
+def generate(user_input): 
+    return get_response(user_input, [])