Update main.py

main.py

@@ -5,54 +5,53 @@ import streamlit as st
 import anthropic
 from requests import JSONDecodeError
 
-from langchain_community.embeddings import HuggingFaceBgeEmbeddings
+# Updated imports for latest LangChain
+from langchain_huggingface import HuggingFaceEmbeddings
 from langchain_community.vectorstores import SupabaseVectorStore
-from langchain_community.llms import HuggingFaceEndpoint
-from langchain_community.chat_models import ChatOpenAI
+from langchain_openai import ChatOpenAI
+from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
 
-from langchain.chains import ConversationalRetrievalChain
+# Updated memory and chain imports
 from langchain.memory import ConversationBufferMemory
+from langchain.chains import create_retrieval_chain
+from langchain.chains.combine_documents import create_stuff_documents_chain
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+from langchain_core.messages import HumanMessage, AIMessage
 
 from supabase import Client, create_client
 from streamlit.logger import get_logger
 from stats import get_usage, add_usage
 
 # ─────── supabase + secrets ────────────────────────────────────────────────────
-supabase_url    = st.secrets.SUPABASE_URL
-supabase_key    = st.secrets.SUPABASE_KEY
-openai_api_key  = st.secrets.openai_api_key
+supabase_url = st.secrets.SUPABASE_URL
+supabase_key = st.secrets.SUPABASE_KEY
+openai_api_key = st.secrets.openai_api_key
 anthropic_api_key = st.secrets.anthropic_api_key
-hf_api_key      = st.secrets.hf_api_key
-username        = st.secrets.username
+hf_api_key = st.secrets.hf_api_key
+username = st.secrets.username
 
 supabase: Client = create_client(supabase_url, supabase_key)
 logger = get_logger(__name__)
 
-# ─────── embeddings ─────────────────────────────────────────────────────────────
-# Switch to local BGE embeddings (no JSONDecode errors, no HTTP‑batch issues) :contentReference[oaicite:0]{index=0}
-embeddings = HuggingFaceBgeEmbeddings(
+# ─────── embeddings (Updated to use langchain-huggingface) ─────────────────────
+embeddings = HuggingFaceEmbeddings(
     model_name="BAAI/bge-large-en-v1.5",
     model_kwargs={"device": "cpu"},
     encode_kwargs={"normalize_embeddings": True}
 )
-# ─────── vector store + memory ─────────────────────────────────────────────────
+
+# ─────── vector store ──────────────────────────────────────────────────────────
 vector_store = SupabaseVectorStore(
     client=supabase,
     embedding=embeddings,
     query_name="match_documents",
     table_name="documents",
 )
-memory = ConversationBufferMemory(
-    memory_key="chat_history",
-    input_key="question",
-    output_key="answer",
-    return_messages=True,
-)
 
 # ─────── LLM setup ──────────────────────────────────────────────────────────────
-model        = "HuggingFaceTB/SmolLM3-3B"
-temperature  = 0.1
-max_tokens   = 500
+model = "HuggingFaceTB/SmolLM3-3B"
+temperature = 0.1
+max_tokens = 500
 
 import re
 
@@ -66,10 +65,10 @@ def clean_response(answer: str) -> str:
     answer = re.sub(r'<thinking>.*?</thinking>', '', answer, flags=re.DOTALL)
     
     # Remove other common AI response artifacts
-    answer = re.sub(r'\[.*?\]', '', answer, flags=re.DOTALL)  # Remove bracketed content
-    answer = re.sub(r'\{.*?\}', '', answer, flags=re.DOTALL)  # Remove curly bracketed content
-    answer = re.sub(r'```.*?```', '', answer, flags=re.DOTALL)  # Remove code blocks
-    answer = re.sub(r'---.*?---', '', answer, flags=re.DOTALL)  # Remove dashed sections
+    answer = re.sub(r'\[.*?\]', '', answer, flags=re.DOTALL)
+    answer = re.sub(r'\{.*?\}', '', answer, flags=re.DOTALL)
+    answer = re.sub(r'```.*?```', '', answer, flags=re.DOTALL)
+    answer = re.sub(r'---.*?---', '', answer, flags=re.DOTALL)
     
     # Remove excessive whitespace and newlines
     answer = re.sub(r'\s+', ' ', answer).strip()
@@ -79,65 +78,93 @@ def clean_response(answer: str) -> str:
     answer = re.sub(r'\s*(Sincerely,.*|Best regards,.*|Regards,.*)$', '', answer, flags=re.IGNORECASE)
     
     return answer
+
+def create_conversational_rag_chain():
+    """Create a modern conversational RAG chain using LCEL."""
+    
+    # Create the HuggingFace LLM
+    llm = ChatOpenAI(
+        base_url=f"https://router.huggingface.co/hf-inference/models/{model}/v1",
+        api_key=hf_api_key,
+        model=model,
+        temperature=temperature,
+        max_tokens=max_tokens,
+        timeout=30,
+        max_retries=3,
+    )
+    
+    # Create retriever
+    retriever = vector_store.as_retriever(
+        search_kwargs={"score_threshold": 0.6, "k": 4, "filter": {"user": username}}
+    )
+    
+    # Create system prompt for RAG
+    system_prompt = """You are a helpful safety assistant. Use the following pieces of retrieved context to answer the question. 
+    If you don't know the answer based on the context, just say that you don't have enough information to answer that question.
+    
+    Context: {context}
+    
+    Chat History: {chat_history}
+    
+    Question: {input}
+    
+    Answer:"""
+    
+    prompt = ChatPromptTemplate.from_messages([
+        ("system", system_prompt),
+        MessagesPlaceholder("chat_history"),
+        ("human", "{input}"),
+    ])
     
-def response_generator(query: str) -> str:
+    # Create document processing chain
+    question_answer_chain = create_stuff_documents_chain(llm, prompt)
+    
+    # Create retrieval chain
+    rag_chain = create_retrieval_chain(retriever, question_answer_chain)
+    
+    return rag_chain
+
+def response_generator(query: str, chat_history: list) -> str:
     """Ask the RAG chain to answer `query`, with JSON‑error fallback."""
     # log usage
     add_usage(supabase, "chat", "prompt:" + query, {"model": model, "temperature": temperature})
     logger.info("Using HF model %s", model)
 
-    # prepare HF text-generation LLM
-    # hf = HuggingFaceEndpoint(
-    #     # endpoint_url=f"https://api-inference.huggingface.co/models/{model}",
-    #     endpoint_url=f"https://router.huggingface.co/hf-inference/models/{model}",
-    #     task="text-generation",
-    #     huggingfacehub_api_token=hf_api_key,
-    #     model_kwargs={
-    #         "temperature": temperature,
-    #         "max_new_tokens": max_tokens,
-    #         "return_full_text": False,
-    #     },
-    # )
-
-    hf = ChatOpenAI(
-            base_url=f"https://router.huggingface.co/hf-inference/models/{model}/v1",
-            api_key=hf_api_key,
-            model=model,
-            temperature=temperature,
-            max_tokens=max_tokens,
-            timeout=30,  # Add timeout
-            max_retries=3,  # Built-in retry logic
-        )
-
-    # conversational RAG chain
-    qa = ConversationalRetrievalChain.from_llm(
-        llm=hf,
-        retriever=vector_store.as_retriever(
-            search_kwargs={"score_threshold": 0.6, "k": 4, "filter": {"user": username}}
-        ),
-        memory=memory,
-        verbose=True,
-        return_source_documents=True,
-    )
+    # Create the RAG chain
+    rag_chain = create_conversational_rag_chain()
+    
+    # Format chat history for the chain
+    formatted_history = []
+    for msg in chat_history:
+        if msg["role"] == "user":
+            formatted_history.append(HumanMessage(content=msg["content"]))
+        elif msg["role"] == "assistant":
+            formatted_history.append(AIMessage(content=msg["content"]))
 
     try:
-        result = qa({"question": query})
+        result = rag_chain.invoke({
+            "input": query,
+            "chat_history": formatted_history
+        })
+        
+        answer = result.get("answer", "")
+        context = result.get("context", [])
+        
+        if not context:
+            return (
+                "I'm sorry, I don't have enough information to answer that. "
+                "If you have a public data source to add, please email [email protected]."
+            )
+
+        answer = clean_response(answer)
+        return answer
+        
     except JSONDecodeError as e:
-        # fallback logging  
-        logger.error("Embedding JSONDecodeError: %s", e)
-        return "Sorry, I had trouble understanding the embedded data. Please try again."
-
-    answer = result.get("answer", "")
-    sources = result.get("source_documents", [])
-
-    if not sources:
-        return (
-            "I’m sorry, I don’t have enough information to answer that. "
-            "If you have a public data source to add, please email [email protected]."
-        )
-
-    answer = clean_response(answer)
-    return answer
+        logger.error("JSONDecodeError: %s", e)
+        return "Sorry, I had trouble processing your request. Please try again."
+    except Exception as e:
+        logger.error("Unexpected error: %s", e)
+        return "Sorry, I encountered an error while processing your request. Please try again."
 
 # ─────── Streamlit UI ──────────────────────────────────────────────────────────
 st.set_page_config(
@@ -161,23 +188,30 @@ st.markdown(
     "|[paper](https://securade.ai/assets/pdfs/Securade.ai-Safety-Copilot-Whitepaper.pdf)]"
 )
 
+# Initialize chat history
 if "chat_history" not in st.session_state:
     st.session_state.chat_history = []
 
-# show history
+# Display chat history
 for msg in st.session_state.chat_history:
     with st.chat_message(msg["role"]):
         st.markdown(msg["content"])
 
-# new user input
+# Handle new user input
 if prompt := st.chat_input("Ask a question"):
+    # Add user message to history
     st.session_state.chat_history.append({"role": "user", "content": prompt})
+    
+    # Display user message
     with st.chat_message("user"):
         st.markdown(prompt)
 
+    # Generate and display response
     with st.spinner("Safety briefing in progress..."):
-        answer = response_generator(prompt)
+        answer = response_generator(prompt, st.session_state.chat_history[:-1])  # Exclude current message
 
     with st.chat_message("assistant"):
         st.markdown(answer)
-    st.session_state.chat_history.append({"role": "assistant", "content": answer})
+    
+    # Add assistant response to history
+    st.session_state.chat_history.append({"role": "assistant", "content": answer})