pipeline.py

import os
import getpass
from pydantic_ai import Agent  # Import the Agent from pydantic_ai
from pydantic_ai.models.mistral import MistralModel  # Import the Mistral model
import spacy  # Import spaCy for NER functionality
import pandas as pd
from typing import Optional
from langchain.docstore.document import Document
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS
from langchain.chains import RetrievalQA
from smolagents import CodeAgent, DuckDuckGoSearchTool, ManagedAgent, LiteLLMModel
import subprocess  # Import subprocess to run shell commands
from langchain.llms.base import LLM  # Import LLM

# Initialize Mistral agent using Pydantic AI
mistral_api_key = os.environ.get("MISTRAL_API_KEY")  # Ensure your Mistral API key is set
mistral_model = MistralModel("mistral-large-latest", api_key=mistral_api_key)  # Use a Mistral model
mistral_agent = Agent(mistral_model)

# Load spaCy model for NER and download the spaCy model if not already installed
def install_spacy_model():
    try:
        spacy.load("en_core_web_sm")
        print("spaCy model 'en_core_web_sm' is already installed.")
    except OSError:
        print("Downloading spaCy model 'en_core_web_sm'...")
        subprocess.run(["python", "-m", "spacy", "download", "en_core_web_sm"], check=True)
        print("spaCy model 'en_core_web_sm' downloaded successfully.")

# Call the function to install the spaCy model if needed
install_spacy_model()

# Load the spaCy model globally
nlp = spacy.load("en_core_web_sm")

# Function to moderate text using Pydantic AI's Mistral moderation model
def moderate_text(query: str) -> str:
    """
    Classifies the query as harmful or not using Mistral Moderation via Pydantic AI.
    Returns "OutOfScope" if harmful, otherwise returns the original query.
    """
    response = mistral_agent.call("classify", {"inputs": [query]})
    categories = response['results'][0]['categories']

    # Check if harmful content is flagged in moderation categories
    if categories.get("violence_and_threats", False) or \
       categories.get("hate_and_discrimination", False) or \
       categories.get("dangerous_and_criminal_content", False) or \
       categories.get("selfharm", False):
        return "OutOfScope"
    return query

# 3) build_or_load_vectorstore (no changes)
def build_or_load_vectorstore(csv_path: str, store_dir: str) -> FAISS:
    if os.path.exists(store_dir):
        print(f"DEBUG: Found existing FAISS store at '{store_dir}'. Loading...")
        embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/multi-qa-mpnet-base-dot-v1")
        vectorstore = FAISS.load_local(store_dir, embeddings)
        return vectorstore
    else:
        print(f"DEBUG: Building new store from CSV: {csv_path}")
        df = pd.read_csv(csv_path)
        df = df.loc[:, ~df.columns.str.contains('^Unnamed')]
        df.columns = df.columns.str.strip()
        if "Answer" in df.columns:
            df.rename(columns={"Answer": "Answers"}, inplace=True)
        if "Question" not in df.columns and "Question " in df.columns:
            df.rename(columns={"Question ": "Question"}, inplace=True)
        if "Question" not in df.columns or "Answers" not in df.columns:
            raise ValueError("CSV must have 'Question' and 'Answers' columns.")
        docs = []
        for _, row in df.iterrows():
            q = str(row["Question"])
            ans = str(row["Answers"])
            doc = Document(page_content=ans, metadata={"question": q})
            docs.append(doc)
        embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/multi-qa-mpnet-base-dot-v1")
        vectorstore = FAISS.from_documents(docs, embedding=embeddings)
        vectorstore.save_local(store_dir)
        return vectorstore

# 4) Build RAG chain for Gemini (no changes)
def build_rag_chain(llm_model: LiteLLMModel, vectorstore: FAISS) -> RetrievalQA:
    class GeminiLangChainLLM(LLM):
        def _call(self, prompt: str, stop: Optional[list] = None, **kwargs) -> str:
            messages = [{"role": "user", "content": prompt}]
            return llm_model(messages, stop_sequences=stop)

        @property
        def _llm_type(self) -> str:
            return "custom_gemini"
    
    retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 3})
    gemini_as_llm = GeminiLangChainLLM()
    rag_chain = RetrievalQA.from_chain_type(
        llm=gemini_as_llm,
        chain_type="stuff",
        retriever=retriever,
        return_source_documents=True
    )
    return rag_chain

# 5) Initialize all the separate chains
classification_chain = get_classification_chain()
refusal_chain = get_refusal_chain()  # Refusal chain will now use dynamic topic
tailor_chain = get_tailor_chain()
cleaner_chain = get_cleaner_chain()

# 6) Build our vectorstores + RAG chains
wellness_csv = "AIChatbot.csv"
brand_csv = "BrandAI.csv"
wellness_store_dir = "faiss_wellness_store"
brand_store_dir = "faiss_brand_store"

wellness_vectorstore = build_or_load_vectorstore(wellness_csv, wellness_store_dir)
brand_vectorstore = build_or_load_vectorstore(brand_csv, brand_store_dir)

gemini_llm = LiteLLMModel(model_id="gemini/gemini-pro", api_key=os.environ.get("GEMINI_API_KEY"))
wellness_rag_chain = build_rag_chain(gemini_llm, wellness_vectorstore)
brand_rag_chain = build_rag_chain(gemini_llm, brand_vectorstore)

# 7) Tools / Agents for web search (no changes)
search_tool = DuckDuckGoSearchTool()
web_agent = CodeAgent(tools=[search_tool], model=gemini_llm)
managed_web_agent = ManagedAgent(agent=web_agent, name="web_search", description="Runs web search for you.")
manager_agent = CodeAgent(tools=[], model=gemini_llm, managed_agents=[managed_web_agent])

def do_web_search(query: str) -> str:
    print("DEBUG: Attempting web search for more info...")
    search_query = f"Give me relevant info: {query}"
    response = manager_agent.run(search_query)
    return response

# 8) Orchestrator: run_with_chain
def run_with_chain(query: str) -> str:
    print("DEBUG: Starting run_with_chain...")
    
    # 1) Moderate the query for harmful content
    moderated_query = moderate_text(query)
    if moderated_query == "OutOfScope":
        return "Sorry, this query contains harmful or inappropriate content."

    # 2) Classify the query
    class_result = classification_chain.invoke({"query": moderated_query})
    classification = class_result.get("text", "").strip()
    print("DEBUG: Classification =>", classification)

    # If OutOfScope => refusal => tailor => return
    if classification == "OutOfScope":
        # Extract the main topic for the refusal message
        topic = extract_main_topic(moderated_query)
        print("DEBUG: Extracted Topic =>", topic)
        
        # Pass the extracted topic to the refusal chain
        refusal_text = refusal_chain.run({"topic": topic})
        final_refusal = tailor_chain.run({"response": refusal_text})
        return final_refusal.strip()

    # If Wellness => wellness RAG => if insufficient => web => unify => tailor
    if classification == "Wellness":
        rag_result = wellness_rag_chain({"query": moderated_query})
        csv_answer = rag_result["result"].strip()
        if not csv_answer:
            web_answer = do_web_search(moderated_query)
        else:
            lower_ans = csv_answer.lower()
            if any(phrase in lower_ans for phrase in ["i do not know", "not sure", "no context", "cannot answer"]):
                web_answer = do_web_search(moderated_query)
            else:
                web_answer = ""
        final_merged = cleaner_chain.merge(kb=csv_answer, web=web_answer)
        final_answer = tailor_chain.run({"response": final_merged})
        return final_answer.strip()

    # If Brand => brand RAG => tailor => return
    if classification == "Brand":
        rag_result = brand_rag_chain({"query": moderated_query})
        csv_answer = rag_result["result"].strip()
        final_merged = cleaner_chain.merge(kb=csv_answer, web="")
        final_answer = tailor_chain.run({"response": final_merged})
        return final_answer.strip()

    # fallback
    refusal_text = refusal_chain.run({"topic": "this topic"})
    final_refusal = tailor_chain.run({"response": refusal_text})
    return final_refusal.strip()