app.py

# Import necessary libraries
import os  # Interacting with the operating system (reading/writing files)
import chromadb  # High-performance vector database for storing/querying dense vectors
from dotenv import load_dotenv  # Loading environment variables from a .env file
import json  # Parsing and handling JSON data

# LangChain imports
from langchain_core.documents import Document  # Document data structures
from langchain_core.runnables import RunnablePassthrough  # LangChain core library for running pipelines
from langchain_core.output_parsers import StrOutputParser  # String output parser
from langchain.prompts import ChatPromptTemplate  # Template for chat prompts
from langchain.chains.query_constructor.base import AttributeInfo  # Base classes for query construction
from langchain.retrievers.self_query.base import SelfQueryRetriever  # Base classes for self-querying retrievers
from langchain.retrievers.document_compressors import LLMChainExtractor, CrossEncoderReranker  # Document compressors
from langchain.retrievers import ContextualCompressionRetriever  # Contextual compression retrievers

# LangChain community & experimental imports
from langchain_community.vectorstores import Chroma  # Implementations of vector stores like Chroma
from langchain_community.document_loaders import PyPDFDirectoryLoader, PyPDFLoader  # Document loaders for PDFs
from langchain_community.cross_encoders import HuggingFaceCrossEncoder  # Cross-encoders from HuggingFace
from langchain_experimental.text_splitter import SemanticChunker  # Experimental text splitting methods
from langchain.text_splitter import (
    CharacterTextSplitter,  # Splitting text by characters
    RecursiveCharacterTextSplitter  # Recursive splitting of text by characters
)
from langchain_core.tools import tool
from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

# LangChain OpenAI imports
from langchain_openai import AzureOpenAIEmbeddings, AzureChatOpenAI  # OpenAI embeddings and models
from langchain.embeddings.openai import OpenAIEmbeddings  # OpenAI embeddings for text vectors

# LlamaParse & LlamaIndex imports
from llama_parse import LlamaParse  # Document parsing library
from llama_index.core import Settings, SimpleDirectoryReader  # Core functionalities of the LlamaIndex

# LangGraph import
from langgraph.graph import StateGraph, END, START  # State graph for managing states in LangChain

# Pydantic import
from pydantic import BaseModel  # Pydantic for data validation

# Typing imports
from typing import Dict, List, Tuple, Any, TypedDict  # Python typing for function annotations

# Other utilities
import numpy as np  # Numpy for numerical operations
from groq import Groq
from mem0 import MemoryClient
import streamlit as st
from datetime import datetime

#====================================SETUP=====================================#
# Fetch secrets from Hugging Face Spaces
api_key = config.get("API_KEY")
endpoint = config.get("OPENAI_API_BASE")
llama_api_key = os.environ['GROQ_API_KEY']
MEM0_api_key = os.environ['MEM0_API_KEY']

# Initialize the OpenAI embedding function for Chroma
embedding_function = chromadb.utils.embedding_functions.OpenAIEmbeddingFunction(
    api_base=endpoint,  # API base endpoint
    api_key=api_key,    # API key
    model_name='text-embedding-ada-002'  # This is a fixed value and does not need modification
)

# This initializes the OpenAI embedding function for the Chroma vectorstore, using the provided endpoint and API key.

# Initialize the OpenAI Embeddings
embedding_model = OpenAIEmbeddings(
    openai_api_base=endpoint,
    openai_api_key=api_key,
    model='text-embedding-ada-002'
)

# Initialize the Chat OpenAI model
llm = ChatOpenAI(
    base_url=endpoint,
    openai_api_key=api_key,
    model="gpt-4o-mini",
    streaming=False
)
# This initializes the Chat OpenAI model with the provided endpoint, API key, deployment name, and a temperature setting of 0 (to control response variability).

# set the LLM and embedding model in the LlamaIndex settings.
Settings.llm = llm          # Define the LLM model
Settings.embedding = embedding_model  # Define the embedding model

#================================Creating Langgraph agent======================#

class AgentState(TypedDict):
    query: str  # The current user query
    expanded_query: str  # The expanded version of the user query
    context: List[Dict[str, Any]]  # Retrieved documents (content and metadata)
    response: str  # The generated response to the user query
    precision_score: float  # The precision score of the response
    groundedness_score: float  # The groundedness score of the response
    groundedness_loop_count: int  # Counter for groundedness refinement loops
    precision_loop_count: int  # Counter for precision refinement loops
    feedback: str
    query_feedback: str
    groundedness_check: bool
    loop_max_iter: int

def expand_query(state):
    """
    Expands the user query to improve retrieval of nutrition disorder-related information.
    """
    print("---------Expanding Query---------")
    system_message = '''You are an assistant that expands user queries about nutrition
disorders to improve information retrieval. Preserve the original intent,
add relevant medical and nutritional terminology, and avoid introducing
new unrelated topics.'''

    expand_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Expand this query: {query} using the feedback: {query_feedback}")
    ])

    chain = expand_prompt | llm | StrOutputParser()
    expanded_query = chain.invoke({"query": state['query'], "query_feedback": state["query_feedback"]})
    print("expanded_query", expanded_query)
    state["expanded_query"] = expanded_query
    return state


# Initialize the Chroma vector store for retrieving documents
vector_store = Chroma(
    collection_name="nutritional_hypotheticals",
    persist_directory="./nutritional_db",
    embedding_function=embedding_model
)

# Create a retriever from the vector store
retriever = vector_store.as_retriever(
    search_type='similarity',
    search_kwargs={'k': 3}
)

def retrieve_context(state):
    """
    Retrieves context from the vector store using the expanded or original query.
    """
    print("---------retrieve_context---------")
    query = state['expanded_query']  # Use the expanded query
    # Retrieve documents from the vector store
    docs = retriever.invoke(query)
    print("Retrieved documents:", docs)  # Debugging: Print the raw docs object

    # Extract both page_content and metadata from each document
    context = [
        {
            "content": doc.page_content,  # The actual content of the document
            "metadata": doc.metadata      # The metadata (e.g., source, page number, etc.)
        }
        for doc in docs
    ]
    state['context'] = context  # Store the context
    print("Extracted context with metadata:", context)  # Debugging
    return state


def craft_response(state: Dict) -> Dict:
    """
    Generates a response using the retrieved context, focusing on nutrition disorders.
    """
    print("---------craft_response---------")
    system_message = '''You are a Nutrition Disorder Specialist. Using only the provided
context from trusted clinical and nutritional references, answer the
user's query in a clear, concise, and clinically accurate way. If the
context does not contain enough information, say so explicitly. Do not
hallucinate or invent facts outside the context.'''

    response_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Query: {query}\nContext: {context}\n\nfeedback: {feedback}")
    ])

    chain = response_prompt | llm
    response_msg = chain.invoke({
        "query": state['query'],
        "context": "\n".join([doc["content"] for doc in state['context']]),
        "feedback": state["feedback"]  # add feedback to the prompt
    })
    # Store only the text content for downstream steps
    state['response'] = response_msg.content
    print("intermediate response: ", response_msg.content)

    return state


def score_groundedness(state: Dict) -> Dict:
    """
    Checks whether the response is grounded in the retrieved context.
    """
    print("---------check_groundedness---------")
    system_message = '''You are an evaluator. Given a context and an answer, assign a single
numeric groundedness score between 0 and 1. 
- 1 means the answer is fully supported by the context.
- 0 means the answer is not supported at all.
Return ONLY the number.'''

    groundedness_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Context: {context}\nResponse: {response}\n\nGroundedness score:")
    ])

    chain = groundedness_prompt | llm | StrOutputParser()
    groundedness_score = float(chain.invoke({
        "context": "\n".join([doc["content"] for doc in state['context']]),
        "response": state['response']  # Use the stored response
    }))
    print("groundedness_score: ", groundedness_score)
    state['groundedness_loop_count'] += 1
    print("#########Groundedness Incremented###########")
    state['groundedness_score'] = groundedness_score

    return state


def check_precision(state: Dict) -> Dict:
    """
    Checks whether the response precisely addresses the user’s query.
    """
    print("---------check_precision---------")
    system_message = '''You are an evaluator. Given a user query and an answer, assign a
precision score between 0 and 1 indicating how directly and completely
the answer addresses the query. 
Return ONLY the number.'''

    precision_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Query: {query}\nResponse: {response}\n\nPrecision score:")
    ])

    chain = precision_prompt | llm | StrOutputParser()
    precision_score = float(chain.invoke({
        "query": state['query'],
        "response": state['response']  # Use the response from the state
    }))
    state['precision_score'] = precision_score
    print("precision_score:", precision_score)
    state['precision_loop_count'] += 1
    print("#########Precision Incremented###########")
    return state


def refine_response(state: Dict) -> Dict:
    """
    Suggests improvements for the generated response.
    """
    print("---------refine_response---------")

    system_message = '''You are a senior clinical nutrition expert. Review the given query and
response. Suggest specific, actionable improvements to make the response
more accurate, complete, and clinically useful.'''

    refine_response_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Query: {query}\nResponse: {response}\n\n"
                 "What improvements can be made to enhance accuracy and completeness?")
    ])

    chain = refine_response_prompt | llm | StrOutputParser()

    # Store response suggestions in a structured format
    feedback = (
        f"Previous Response: {state['response']}\nSuggestions: "
        f"{chain.invoke({'query': state['query'], 'response': state['response']})}"
    )
    print("feedback: ", feedback)
    print(f"State: {state}")
    state['feedback'] = feedback
    return state


def refine_query(state: Dict) -> Dict:
    """
    Suggests improvements for the expanded query.
    """
    print("---------refine_query---------")
    system_message = '''You are an expert query engineer. Given the original query and the
current expanded query, suggest how to improve the expanded query to
retrieve more precise and relevant information about nutrition
disorders. Return only the improved expanded query.'''

    refine_query_prompt = ChatPromptTemplate.from_messages([
        ("system", system_message),
        ("user", "Original Query: {query}\nExpanded Query: {expanded_query}\n\n"
                 "What improvements can be made for a better search?")
    ])

    chain = refine_query_prompt | llm | StrOutputParser()

    # Store refinement suggestions without modifying the original expanded query
    query_feedback = (
        f"Previous Expanded Query: {state['expanded_query']}\nSuggestions: "
        f"{chain.invoke({'query': state['query'], 'expanded_query': state['expanded_query']})}"
    )
    print("query_feedback: ", query_feedback)
    print(f"Groundedness loop count: {state['groundedness_loop_count']}")
    state['query_feedback'] = query_feedback
    return state


def should_continue_groundedness(state):
    """Decides if groundedness is sufficient or needs improvement."""
    print("---------should_continue_groundedness---------")
    print("groundedness loop count: ", state['groundedness_loop_count'])
    if state['groundedness_score'] >= 0.7:  # Threshold for groundedness
        print("Moving to precision")
        return "check_precision"
    else:
        if state["groundedness_loop_count"] > state['loop_max_iter']:
            return "max_iterations_reached"
        else:
            print("---------Groundedness Score Threshold Not met. Refining Response-----------")
            return "refine_response"


def should_continue_precision(state: Dict) -> str:
    """Decides if precision is sufficient or needs improvement."""
    print("---------should_continue_precision---------")
    print("precision loop count: ", state['precision_loop_count'])
    if state['precision_score'] >= 0.7:  # Threshold for precision
        return "pass"  # Complete the workflow
    else:
        if state["precision_loop_count"] > state['loop_max_iter']:  # Maximum allowed loops
            return "max_iterations_reached"
        else:
            print("---------Precision Score Threshold Not met. Refining Query-----------")  # Debugging
            return "refine_query"  # Refine the query


def max_iterations_reached(state: Dict) -> Dict:
    """Handles the case when the maximum number of iterations is reached."""
    print("---------max_iterations_reached---------")
    response = "I'm unable to refine the response further. Please provide more context or clarify your question."
    state['response'] = response
    return state


from langgraph.graph import END, StateGraph, START

def create_workflow() -> StateGraph:
    """Creates the updated workflow for the AI nutrition agent."""
    workflow = StateGraph(AgentState)  # Initial state type

    # Add processing nodes
    workflow.add_node("expand_query", expand_query)         # Step 1
    workflow.add_node("retrieve_context", retrieve_context) # Step 2
    workflow.add_node("craft_response", craft_response)     # Step 3
    workflow.add_node("score_groundedness", score_groundedness)  # Step 4
    workflow.add_node("refine_response", refine_response)   # Step 5
    workflow.add_node("check_precision", check_precision)   # Step 6
    workflow.add_node("refine_query", refine_query)         # Step 7
    workflow.add_node("max_iterations_reached", max_iterations_reached)  # Step 8

    # Main flow edges
    workflow.add_edge(START, "expand_query")
    workflow.add_edge("expand_query", "retrieve_context")
    workflow.add_edge("retrieve_context", "craft_response")
    workflow.add_edge("craft_response", "score_groundedness")

    # Conditional edges based on groundedness check
    workflow.add_conditional_edges(
        "score_groundedness",
        should_continue_groundedness,  # Use the conditional function
        {
            "check_precision": "check_precision",      # If well-grounded, proceed to precision check.
            "refine_response": "refine_response",      # If not, refine the response.
            "max_iterations_reached": "max_iterations_reached"  # If max loops reached, exit.
        }
    )

    workflow.add_edge("refine_response", "craft_response")  # Refined responses are reprocessed.

    # Conditional edges based on precision check
    workflow.add_conditional_edges(
        "check_precision",
        should_continue_precision,  # Use the conditional function
        {
            "pass": END,                    # If precise, complete the workflow.
            "refine_query": "refine_query", # If imprecise, refine the query.
            "max_iterations_reached": "max_iterations_reached"  # If max loops reached, exit.
        }
    )

    workflow.add_edge("refine_query", "expand_query")  # Refined queries go through expansion again.

    workflow.add_edge("max_iterations_reached", END)

    return workflow


#=========================== Defining the agentic rag tool ====================#
WORKFLOW_APP = create_workflow().compile()

@tool
def agentic_rag(query: str):
    """
    Runs the RAG-based agent with conversation history for context-aware responses.
    """
    # Initialize state with necessary parameters
    inputs = {
        "query": query,        # Current user query
        "expanded_query": "",  # Expanded version (to be filled by expand_query)
        "context": [],         # Retrieved documents (initially empty)
        "response": "",        # AI-generated response
        "precision_score": 0.0,
        "groundedness_score": 0.0,
        "groundedness_loop_count": 0,
        "precision_loop_count": 0,
        "feedback": "",
        "query_feedback": "",
        "loop_max_iter": 3     # Maximum number of iterations for loops
    }

    output = WORKFLOW_APP.invoke(inputs)

    return output


#================================ Guardrails ===========================#
llama_guard_client = Groq(api_key=llama_api_key)

# Function to filter user input with Llama Guard
def filter_input_with_llama_guard(user_input, model="meta-llama/llama-guard-4-12b"):
    """
    Filters user input using Llama Guard to ensure it is safe.
    """
    try:
        response = llama_guard_client.chat.completions.create(
            messages=[{"role": "user", "content": user_input}],
            model=model,
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        print(f"Error with Llama Guard: {e}")
        return None


#============================= Adding Memory to the agent using mem0 ===============================#

class NutritionBot:
    def __init__(self):
        """
        Initialize the NutritionBot class, setting up memory, the LLM client, tools, and the agent executor.
        """

        # Initialize a memory client to store and retrieve customer interactions
        self.memory = MemoryClient(api_key=MEM0_api_key)

        # Initialize the OpenAI client using the provided credentials
        self.client = ChatOpenAI(
            model="gpt-4o-mini",              # Model to use
            openai_api_key=config.get("API_KEY"),  # API key for authentication
            base_url=config.get("OPENAI_API_BASE"),
            temperature=0                     # Deterministic responses
        )

        # Define tools available to the chatbot, such as web search
        tools = [agentic_rag]

        # Define the system prompt to set the behavior of the chatbot
        system_prompt = """You are a caring and knowledgeable Medical Support Agent, specializing in nutrition disorder-related guidance. Your goal is to provide accurate, empathetic, and tailored nutritional recommendations while ensuring a seamless customer experience.
                          Guidelines for Interaction:
                          Maintain a polite, professional, and reassuring tone.
                          Show genuine empathy for customer concerns and health challenges.
                          Reference past interactions to provide personalized and consistent advice.
                          Engage with the customer by asking about their food preferences, dietary restrictions, and lifestyle before offering recommendations.
                          Ensure consistent and accurate information across conversations.
                          If any detail is unclear or missing, proactively ask for clarification.
                          Always use the agentic_rag tool to retrieve up-to-date and evidence-based nutrition insights.
                          Keep track of ongoing issues and follow-ups to ensure continuity in support.
                          Your primary goal is to help customers make informed nutrition decisions that align with their health conditions and personal preferences.
        """

        # Build the prompt template for the agent
        prompt = ChatPromptTemplate.from_messages([
            ("system", system_prompt),        # System instructions
            ("human", "{input}"),             # Placeholder for human input
            ("placeholder", "{agent_scratchpad}")  # Intermediate reasoning steps
        ])

        # Create an agent capable of interacting with tools and executing tasks
        agent = create_tool_calling_agent(self.client, tools, prompt)

        # Wrap the agent in an executor to manage tool interactions and execution flow
        self.agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

    def store_customer_interaction(self, user_id: str, message: str, response: str, metadata: Dict = None):
        """
        Store customer interaction in memory for future reference.
        """
        if metadata is None:
            metadata = {}

        # Add a timestamp to the metadata for tracking purposes
        metadata["timestamp"] = datetime.now().isoformat()

        # Format the conversation for storage
        conversation = [
            {"role": "user", "content": message},
            {"role": "assistant", "content": response}
        ]

        # Store the interaction in the memory client
        self.memory.add(
            conversation,
            user_id=user_id,
            output_format="v1.1",
            metadata=metadata
        )

    def get_relevant_history(self, user_id: str, query: str) -> List[Dict]:
        """
        Retrieve past interactions relevant to the current query.
        """
        return self.memory.search(
            query=query,    # Search for interactions related to the query
            user_id=user_id,
            limit=5         # Limit for retrieved interactions
        )

    def handle_customer_query(self, user_id: str, query: str) -> str:
        """
        Process a customer's query and provide a response, taking into account past interactions.
        """

        # Retrieve relevant past interactions for context
        relevant_history = self.get_relevant_history(user_id, query)

        # Build a context string from the relevant history
        context = "Previous relevant interactions:\n"
        for memory in relevant_history:
            context += f"Memory: {memory['memory']}\n"
            context += "---\n"

        # Print context for debugging purposes
        print("Context: ", context)

        # Prepare a prompt combining past context and the current query
        prompt = f"""
        Context:
        {context}

        Current customer query: {query}

        Provide a helpful response that takes into account any relevant past interactions.
        """

        # Generate a response using the agent
        response = self.agent_executor.invoke({"input": prompt})

        # Store the current interaction for future reference
        self.store_customer_interaction(
            user_id=user_id,
            message=query,
            response=response["output"],
            metadata={"type": "support_query"}
        )

        # Return the chatbot's response
        return response['output']


#=====================User Interface using streamlit ===========================#
def nutrition_disorder_streamlit():
    """
    A Streamlit-based UI for the Nutrition Disorder Specialist Agent.
    """
    st.title("Nutrition Disorder Specialist")
    st.write("Ask me anything about nutrition disorders, symptoms, causes, treatments, and more.")
    st.write("Type 'exit' to end the conversation.")

    # Initialize session state for chat history and user_id if they don't exist
    if 'chat_history' not in st.session_state:
        st.session_state.chat_history = []
    if 'user_id' not in st.session_state:
        st.session_state.user_id = None

    # Login form: Only if user is not logged in
    if st.session_state.user_id is None:
        with st.form("login_form", clear_on_submit=True):
            user_id = st.text_input("Please enter your name to begin:")
            submit_button = st.form_submit_button("Login")
            if submit_button and user_id:
                st.session_state.user_id = user_id
                st.session_state.chat_history.append({
                    "role": "assistant",
                    "content": f"Welcome, {user_id}! How can I help you with nutrition disorders today?"
                })
                st.session_state.login_submitted = True  # Set flag to trigger rerun
        if st.session_state.get("login_submitted", False):
            st.session_state.pop("login_submitted")
            st.rerun()
    else:
        # Display chat history
        for message in st.session_state.chat_history:
            with st.chat_message(message["role"]):
                st.write(message["content"])

        # Chat input with custom placeholder text
        user_query = st.chat_input("Type your question here (or 'exit' to end)...")  # Blank #1
        if user_query:
            if user_query.lower() == "exit":
                st.session_state.chat_history.append({"role": "user", "content": "exit"})
                with st.chat_message("user"):
                    st.write("exit")
                goodbye_msg = "Goodbye! Feel free to return if you have more questions about nutrition disorders."
                st.session_state.chat_history.append({"role": "assistant", "content": goodbye_msg})
                with st.chat_message("assistant"):
                    st.write(goodbye_msg)
                st.session_state.user_id = None
                st.rerun()
                return

            st.session_state.chat_history.append({"role": "user", "content": user_query})
            with st.chat_message("user"):
                st.write(user_query)

            # Filter input using Llama Guard
            filtered_result = filter_input_with_llama_guard(user_query)  # Blank #2
            if filtered_result:
                filtered_result = filtered_result.replace("\n", " ")  # Normalize the result
            else:
                filtered_result = "SAFE"

            # Check if input is safe based on allowed statuses
            if filtered_result in ["SAFE", "S6", "S7"]:  # Blanks #3, #4, #5
                try:
                    if 'chatbot' not in st.session_state:
                        st.session_state.chatbot = NutritionBot()  # Blank #6
                    response = st.session_state.chatbot.handle_customer_query(
                        st.session_state.user_id, user_query
                    )  # Blank #7
                    st.write(response)
                    st.session_state.chat_history.append({"role": "assistant", "content": response})
                except Exception as e:
                    error_msg = f"Sorry, I encountered an error while processing your query. Please try again. Error: {str(e)}"
                    st.write(error_msg)
                    st.session_state.chat_history.append({"role": "assistant", "content": error_msg})
            else:
                inappropriate_msg = "I apologize, but I cannot process that input as it may be inappropriate. Please try again."
                st.write(inappropriate_msg)
                st.session_state.chat_history.append({"role": "assistant", "content": inappropriate_msg})


if __name__ == "__main__":
    nutrition_disorder_streamlit()