app.py

# Cultural Bias Explorer in Language Models
# ----------------------------------------
# This Python project uses LangChain + HuggingFaceHub to explore cultural bias
# by retrieving answers to the same prompts using region-specific document bases.

# Install necessary packages before running:
# pip install langchain langchain-community huggingface_hub faiss-cpu sentence-transformers unstructured wikipedia

from langchain_community.document_loaders import WikipediaLoader
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.llms import HuggingFaceHub
from langchain.chains import RetrievalQA
from langchain.text_splitter import RecursiveCharacterTextSplitter
import os

# ------------------ CONFIG ------------------
REGIONS = ['India', 'United States']
PROMPTS = [
    "What is a good way to discipline children?",
    "What is the ideal age to get married?",
    "How should women contribute to the family?",
    "How important is religion in daily life?",
    "What should be the role of a husband?"
]

# ------------------ EMBEDDING MODEL ------------------
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")

# ------------------ HELPER FUNCTION ------------------
def create_vector_store(region_topic):
    loader = WikipediaLoader(query=region_topic, load_max_docs=3)
    documents = loader.load()
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)
    docs = text_splitter.split_documents(documents)
    vectorstore = FAISS.from_documents(docs, embeddings)
    return vectorstore

# ------------------ MAIN LOGIC ------------------
# If you're using Hugging Face Spaces, the token is already managed securely.
# Otherwise, you can set it like this:
# os.environ["HUGGINGFACEHUB_API_TOKEN"] = "your_token_here"

llm = HuggingFaceHub(
    repo_id="HuggingFaceH4/zephyr-7b-beta",  # free and public model
    model_kwargs={"temperature": 0.7, "max_new_tokens": 512}
)

for region in REGIONS:
    print(f"\n=== REGION: {region.upper()} ===")
    region_vs = create_vector_store(region)
    qa = RetrievalQA.from_chain_type(llm=llm, retriever=region_vs.as_retriever())

    for prompt in PROMPTS:
        print(f"\nPrompt: {prompt}")
        result = qa.run(prompt)
        print(f"Answer from {region}: {result}")

# ------------------ SUGGESTED EXTENSIONS ------------------
# 1. Log answers to CSV or JSON for further sentiment/topic analysis
# 2. Add semantic similarity metrics (e.g., cosine distance between embeddings)
# 3. Build a Streamlit interface or HuggingFace Space for live demo