app.py

# ref: https://github.com/plaban1981/Agents/blob/main/Contextual_Retrieval_processing_prompt.ipynb
import streamlit as st
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema import Document
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS
from langchain.retrievers import BM25Retriever, ContextualCompressionRetriever, EnsembleRetriever
from langchain.retrievers.document_compressors import FlashrankRerank
from langchain_community.document_transformers.embeddings_redundant_filter import EmbeddingsRedundantFilter
from langchain_groq import ChatGroq
from langchain.prompts import ChatPromptTemplate
import hashlib
from typing import List

# Contextual Retrieval Class
class ContextualRetrieval:
    def __init__(self):
        self.text_splitter = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=100)
        model_name = "BAAI/bge-large-en-v1.5"
        self.embeddings = HuggingFaceEmbeddings(model_name=model_name, model_kwargs={'device': 'cpu'})
        self.llm = ChatGroq(model="llama-3.2-3b-preview", temperature=0)

    def process_document(self, document: str) -> List[Document]:
        return self.text_splitter.create_documents([document])

    def generate_contextualized_chunks(self, document: str, chunks: List[Document]) -> List[Document]:
        contextualized_chunks = []
        for chunk in chunks:
            context = self._generate_context(document, chunk.page_content)
            contextualized_content = f"{context}\n\n{chunk.page_content}"
            contextualized_chunks.append(Document(page_content=contextualized_content))
        return contextualized_chunks

    def _generate_context(self, document: str, chunk: str) -> str:
        prompt = ChatPromptTemplate.from_template("""
        Based on the document and a specific chunk of text, generate a 2-3 sentence summary that contextualizes the chunk:
        Document:
        {document}

        Chunk:
        {chunk}

        Context:
        """)
        messages = prompt.format_messages(document=document, chunk=chunk)
        response = self.llm.invoke(messages)
        return response.content.strip()

    def create_vectorstore(self, chunks: List[Document]) -> FAISS:
        return FAISS.from_documents(chunks, self.embeddings)

    def create_bm25_retriever(self, chunks: List[Document]) -> BM25Retriever:
        return BM25Retriever.from_documents(chunks)

    def create_reranker(self, vectorstore):
        retriever = vectorstore.as_retriever(search_kwargs={"k": 10})
        return ContextualCompressionRetriever(base_compressor=FlashrankRerank(), base_retriever=retriever)

    def create_ensemble(self, vectorstore, bm25_retriever):
        vector_retriever = vectorstore.as_retriever(search_kwargs={"k": 10})
        return EnsembleRetriever(
            retrievers=[vector_retriever, bm25_retriever],
            weights=[0.5, 0.5]
        )

    def generate_answer(self, query: str, docs: List[Document]) -> str:
        prompt = ChatPromptTemplate.from_template("""
        Question: {query}
        Relevant Information: {chunks}
        Answer:
        """)
        messages = prompt.format_messages(query=query, chunks="\n\n".join([doc.page_content for doc in docs]))
        response = self.llm.invoke(messages)
        return response.content.strip()

# Streamlit UI
def main():
    st.title("Interactive Ranking and Retrieval Analysis")
    st.write("Experiment with multiple retrieval methods, ranking techniques, and dynamic contextualization.")

    # Document Upload
    uploaded_file = st.file_uploader("Upload a Text Document", type=['txt', 'md'])
    if uploaded_file:
        document = uploaded_file.read().decode("utf-8")
        st.success("Document uploaded successfully!")

        # Initialize Retrieval System
        cr = ContextualRetrieval()
        chunks = cr.process_document(document)
        contextualized_chunks = cr.generate_contextualized_chunks(document, chunks)

        # Create indexes and retrievers
        original_vectorstore = cr.create_vectorstore(chunks)
        contextualized_vectorstore = cr.create_vectorstore(contextualized_chunks)
        original_bm25_retriever = cr.create_bm25_retriever(chunks)
        contextualized_bm25_retriever = cr.create_bm25_retriever(contextualized_chunks)

        # Rerankers and Ensemble Retrievers
        original_reranker = cr.create_reranker(original_vectorstore)
        contextualized_reranker = cr.create_reranker(contextualized_vectorstore)
        original_ensemble = cr.create_ensemble(original_vectorstore, original_bm25_retriever)
        contextualized_ensemble = cr.create_ensemble(contextualized_vectorstore, contextualized_bm25_retriever)

        # Query Input
        query = st.text_input("Enter your query:")
        if query:
            with st.spinner("Fetching results..."):
                # Retrieve results
                original_vector_results = original_vectorstore.similarity_search(query, k=3)
                contextualized_vector_results = contextualized_vectorstore.similarity_search(query, k=3)
                original_bm25_results = original_bm25_retriever.get_relevant_documents(query)
                contextualized_bm25_results = contextualized_bm25_retriever.get_relevant_documents(query)
                original_reranker_results = original_reranker.invoke(query)
                contextualized_reranker_results = contextualized_reranker.invoke(query)
                original_ensemble_results = original_ensemble.invoke(query)
                contextualized_ensemble_results = contextualized_ensemble.invoke(query)

                # Generate answers
                original_vector_answer = cr.generate_answer(query, original_vector_results)
                contextualized_vector_answer = cr.generate_answer(query, contextualized_vector_results)
                original_bm25_answer = cr.generate_answer(query, original_bm25_results)
                contextualized_bm25_answer = cr.generate_answer(query, contextualized_bm25_results)
                original_reranker_answer = cr.generate_answer(query, original_reranker_results)
                contextualized_reranker_answer = cr.generate_answer(query, contextualized_reranker_results)
                original_ensemble_answer = cr.generate_answer(query, original_ensemble_results)
                contextualized_ensemble_answer = cr.generate_answer(query, contextualized_ensemble_results)

            # Display Results
            st.subheader("Results Comparison")

            col1, col2 = st.columns(2)
            with col1:
                st.write("### Original Results")
                st.write("**Vector Search Answer**")
                st.info(original_vector_answer)
                st.write("**BM25 Search Answer**")
                st.info(original_bm25_answer)
                st.write("**Reranker Answer**")
                st.info(original_reranker_answer)
                st.write("**Ensemble Answer**")
                st.info(original_ensemble_answer)

            with col2:
                st.write("### Contextualized Results")
                st.write("**Vector Search Answer**")
                st.info(contextualized_vector_answer)
                st.write("**BM25 Search Answer**")
                st.info(contextualized_bm25_answer)
                st.write("**Reranker Answer**")
                st.info(contextualized_reranker_answer)
                st.write("**Ensemble Answer**")
                st.info(contextualized_ensemble_answer)

if __name__ == "__main__":
    main()